Tag: OpenAI

Better AI Image Prompts: 7 Techniques That Sharpen Your Results

Better AI image prompts follow a five-part formula: subject, style, lighting, mood, detail. These 7 techniques get sharper results from your first generation.

When I typed “a dog in a field” into an AI image generator for the first time, I got back something that looked like stock art from 2008. The tool was fine — my prompt was hollow. Better AI image prompts follow one rule: give the generator enough structure to make intentional choices instead of random ones.

This structure works across DALL-E 3, Adobe Firefly, Canva AI, and most other generators. Most people stop at describing the subject, leaving the generator to guess at style, lighting, and color palette. Give it that direction and your first-try hit rate improves immediately.

Quick Answer

Better AI image prompts include five parts: subject, art style or medium, lighting, mood, and one technical detail like camera angle or lens type. Adding these takes under ten seconds. Generators like DALL-E 3 and Adobe Firefly show an immediate quality improvement when all five are present.

What Goes Into a Stronger AI Image Prompt?

Think of a prompt as a creative brief, not a search query. A search engine tolerates vague keywords; an image generator needs detail to make consistent visual decisions.

I write every prompt using this five-part formula:

  1. Subject — what or who is in the image, with any action described
  2. Style or medium — photograph, oil painting, digital illustration, watercolor
  3. Lighting — golden hour, soft overcast, studio lighting, dramatic side light
  4. Mood or atmosphere — peaceful, tense, nostalgic, futuristic
  5. Technical detail — 85mm lens, shallow depth of field, 35mm film grain, aerial view

Compare “a dog in a field” with “a golden retriever sprinting through a sunlit meadow, photorealistic DSLR photo, golden-hour backlight, joyful atmosphere, 85mm lens.” The second prompt gives the generator five anchors. The output is dramatically different — and consistently so.

Pro Tip

Write the subject first. Most generators weight the beginning of a prompt more than the end. Your most important element should appear in the first five words.

A prompt is a brief, not a keyword — five clear elements give generators enough signal to make consistent choices instead of default-mediocre ones.

How Does Lighting Description Change the Result?

Lighting is the fastest single upgrade to any image prompt. When I added “soft golden-hour backlight” to an otherwise flat prompt, the output shifted from a harsh midday snapshot to something that looked like a film still.

These lighting terms work reliably across most generators:

  • Golden hour / blue hour — warm, directional light; great for outdoor scenes
  • Soft overcast — even natural light; ideal for portraits
  • Studio / three-point lighting — crisp, commercial-grade results
  • Neon glow / cyberpunk ambient — vivid colored light for night scenes
  • Chiaroscuro — high contrast between light and shadow; cinematic and fine-art looks
  • Rim light — subject outlined in light; adds depth and drama

Troubleshooting Tip

If an image looks flat or washed out, the fix is almost always a missing lighting direction. Add “side lighting from the left” or “backlit against a bright window” to any portrait prompt and depth appears immediately.

Lighting direction transforms a flat output into one with depth, mood, and a clear focal point — it is the highest-return addition to any image prompt.

Which Art Style Keywords Actually Work?

Style keywords tell the generator which visual tradition to draw from. “Photograph” and “oil painting” produce completely different images from an identical subject. I tested “a mountain at sunrise” across four style families:

Style Family Prompt Keywords Strongest Generator
Photorealistic DSLR photo, 35mm film, RAW photo, photorealistic DALL-E 3, Firefly
Illustration digital illustration, flat design, vector art, cartoon Canva AI, Ideogram
Fine art oil painting, watercolor, impressionist, charcoal sketch DALL-E 3, Firefly
Concept art cinematic concept art, matte painting, sci-fi artwork DALL-E 3, Ideogram

The impressionist oil painting version was the one I actually used for a real project — it had texture and warmth the photorealistic result lacked for a landscape subject. Style changes the emotional register of an image, not just its appearance. For a full breakdown of which tools handle each style best, see my comparison of free AI image generators: DALL-E 3, Firefly, and Canva AI.

Style keywords act as genre labels — they point the generator toward a specific visual vocabulary and prevent the bland, style-neutral output that appears when style is left out.

How Do Negative Prompts Remove Unwanted Elements?

Negative prompts tell the generator what to leave out. Adobe Firefly has a dedicated negative prompt field; DALL-E 3 accepts exclusions as natural language inside the main prompt.

Elements I exclude from almost every generation:

  • blurry, out of focus, low resolution
  • extra limbs, distorted hands
  • watermark, text overlay, signature
  • oversaturated colors, garish tones, HDR artifacts

In DALL-E 3, I append this sentence to the main prompt: “No watermarks, no text, realistic hand proportions, sharp focus.” In Firefly, these go into the dedicated Negative Prompt box in the generation settings. Adobe’s full documentation is at firefly.adobe.com.

Negative prompts are the eraser — they preemptively remove a generator’s default bad habits before they show up in the output.

Why Does Iterating Beat Starting Over?

My biggest early mistake was discarding every disappointing result and rewriting the entire prompt. That throws away every correct decision the generator made in the first attempt.

Instead, I identify the one element I dislike most and change only that:

  1. Generate the image.
  2. Identify the single biggest problem — lighting, composition, or style.
  3. Change only that element in the prompt, then regenerate.
  4. Repeat until the image matches the intent.

DALL-E 3 and Firefly both support regional editing — painting over a specific area and re-prompting just that region. This preserves composition that was already working. The same habit applies to text: the techniques I use for writing sharper ChatGPT prompts transfer directly to image prompting.

Iterating on one variable at a time turns lucky first attempts into consistent, repeatable results — and it is faster than starting from a blank prompt.

What Are the Most Common AI Image Prompt Mistakes?

  1. Vague subject lines. “A person” is not a prompt. Fix: describe what the subject is doing and where — “a woman reading on a park bench, dappled afternoon sunlight.”
  2. Skipping style entirely. Without a style keyword, generators default to a generic midpoint. Fix: add one word — “photograph,” “watercolor,” or “illustration” — to every prompt.
  3. Using abstract mood words alone. “Dramatic” gives the generator nothing visual to render. Fix: describe mood with physical details — “dark storm clouds, long shadows, cool blue tones.”
  4. Full regeneration instead of regional editing. Don’t discard a good composition because one detail is wrong. Fix: use the inpaint or edit tool to change only the problem area.
  5. Ignoring aspect ratio. Most generators default to square. Fix: set the ratio before generating — 16:9 for banners, 4:5 for social posts, 2:3 for portraits.

Most prompt mistakes share the same root: giving the generator too little to work with, then discarding the whole result rather than fixing the one thing that was wrong.

Frequently Asked Questions

What is the single most important element of an AI image prompt?

The subject with an action or context. Without it, the generator fills the gap randomly. I always start with “who or what is doing what and where” — for example: “a chef plating food in a busy restaurant kitchen at night.”

How long should an AI image prompt be?

Between 20 and 50 words covers most cases. Prompts over 80 words often produce conflicting outputs in DALL-E 3. Cover the five formula elements and stop — extra length rarely improves results.

Do negative prompts work in DALL-E 3?

Yes, as plain language inside the main prompt. I append exclusions at the end: “no text, no watermarks, realistic hand proportions, sharp focus.” This eliminates most hand distortion on the first try.

Why do AI-generated hands still look strange?

Training data historically contained fewer clean close-ups of hands than faces, so generators learned them less reliably. Adding “realistic hand proportions, five fingers, sharp detail” to your prompt reduces the problem significantly. DALL-E 3 handles hands better than most generators today, but the explicit instruction still helps.

Can I reuse the same prompt across different AI image generators?

Yes, with minor adjustments. The five-part structure works everywhere. The key difference is Firefly’s separate Negative Prompt field versus DALL-E 3’s inline exclusions. I test new prompts on both since “photorealistic” triggers different visual outputs in each tool.

Conclusion

Better AI image prompts are not about magic words — they are about giving the generator enough structure to make intentional visual choices. Start with the five-part formula, add a lighting direction, exclude common artifacts, and iterate rather than restart. The improvement shows up on the very next generation.

Ready to pick the right generator to try these on? See my roundup of free AI image tools that skip the paywall and put the formula to work today.

AI Tokens and Context Window Explained: What Every User Needs to Know

AI tokens and context windows explained in plain English — learn the limits that shape every AI conversation and how to work within them effectively.

If you have ever pasted a long document into ChatGPT or Claude and watched the AI forget what you said at the top, you already know what an AI token limit feels like — you just did not have a name for it. Understanding ai tokens and context windows explained in plain terms turns that frustrating quirk into something you can actually predict and work around.

The single most important insight: a context window is the AI’s working memory, and once it fills up, older content does not get summarized — it simply disappears from the model’s view.

Quick Answer

A token is a chunk of text — roughly three to four characters or three-quarters of a word in English. The context window is the total number of tokens an AI can process at once, covering both your input and its reply. When a conversation exceeds that limit, the AI drops the oldest content first.

What Exactly Is an AI Token?

Think of a token as the smallest unit an AI reads. It is not a full word and not a single character — it sits somewhere in between. Most common English words are one token, but longer or unusual words split into several. The word “tokenization,” for example, typically breaks into three tokens: “token,” “ization” is sometimes split further depending on the model.

How Token Counting Works in Practice

I pasted a 100-word email into OpenAI’s free Tokenizer tool and got back 131 tokens — about 1.3 tokens per word, which is typical for English prose. Code and technical content with symbols or non-ASCII characters can run considerably higher, sometimes two tokens per character.

The token count on your AI plan covers both directions: every word you type and every word the model writes back. That combined total is what gets measured against the limit.

Pro tip: To estimate your token count before pasting, multiply your word count by 1.3. A 2,000-word document runs roughly 2,600 tokens — well within most modern context windows, but stack several documents together and it adds up fast.

Tokens are the universal measurement unit AI companies use for both billing and length limits — knowing the rough conversion helps you predict behavior before a session goes sideways.

What Is a Context Window?

The context window is the total number of tokens an AI model can hold in its view at any one moment. It covers the entire conversation: any hidden system prompt the app adds, every message you have sent, and every reply the model has generated. Nothing outside that window is visible to the model — not earlier sessions, not files you shared previously.

Why the Limit Exists

Current AI models process everything inside the context window simultaneously using a technique called attention, which weighs every token against every other token. That computation scales with the square of the token count, which is why a true “infinite” window is not yet practical. Longer windows require significantly more compute and memory per response.

What Happens When You Hit the Limit

When a conversation grows beyond the context window, the app typically drops the oldest messages silently. I noticed this firsthand while editing a long manuscript with Claude — the AI suddenly stopped referencing a character I had introduced 30 exchanges earlier. The character had not changed; the conversation had simply pushed that section out of view.

If you have ever seen ChatGPT cut off mid-answer on a long task, token limits are often the cause. The guide to recovering a full ChatGPT response covers the exact prompts I use to pick up exactly where the model stopped.

The context window is the AI’s working memory: powerful within its boundary, completely blind beyond it.

How Does the Context Window Affect Your Results?

For short tasks — a quick question, a 300-word rewrite — the context window size barely matters. For longer work — editing a 10,000-word report, debugging a large codebase, or running a multi-turn research session — window size is the single biggest factor in whether the AI stays coherent throughout.

Picking the Right Tool for Long Tasks

I check the context window size before starting any task I expect to run long. The Claude AI free plan breakdown shows how the daily limits interact with context length — a useful reference for planning multi-step work on a free tier.

Troubleshooting tip: If the AI starts contradicting an instruction you gave early in the session, the conversation has likely grown past the effective context range. Start a fresh chat and paste in only the background that matters.

Context window size only matters when you are working with large, continuous content — for most everyday tasks, even a 16,000-token window is far more than enough.

How Do Token Limits Compare Across AI Tools?

Context window sizes vary widely across models, and that difference matters the moment your task exceeds a few thousand words. Here is a snapshot of current limits for the most widely used AI tools:

AI Tool Context Window Best Suited For
Gemini 1.5 Pro 1,000,000 tokens (~750,000 words) Very large files, video transcripts
Claude 3.5 Sonnet 200,000 tokens (~150,000 words) Long documents, books, full codebases
ChatGPT-4o 128,000 tokens (~96,000 words) Research, writing, coding sessions
ChatGPT-3.5 (legacy) 16,385 tokens (~12,000 words) Short tasks, quick single-turn questions

Word counts in the table are approximate. Code, tables, and non-English text typically cost more tokens per line than plain English prose.

For a side-by-side look at how two of these tools handle sustained research sessions, the NotebookLM vs ChatGPT research comparison shows exactly where context handling makes a practical difference.

Larger context windows keep the AI coherent over longer work, but they do not eliminate the need to be selective about what you paste — more room just means the wall is farther away, not gone.

Common Mistakes to Avoid

  1. Assuming the AI remembers between sessions. Each new chat starts with a blank context window. Nothing from yesterday is visible. Fix: keep a short “briefing note” with the key facts you repeat across sessions and paste it at the start.
  2. Pasting the entire document when only a section is needed. Flooding the context with irrelevant content leaves less room for the conversation that follows. Fix: paste the relevant excerpt and a one-paragraph summary of the rest.
  3. Confusing the context limit with the output limit. Many models cap both how much you can send and how long a single reply can be — separately. Fix: if the AI stops mid-answer, a simple “continue” prompt usually resumes it.
  4. Ignoring the hidden system prompt. Every AI app prepends a system prompt you never see. On some tools it is thousands of tokens long. Fix: for very long tasks, use a direct API call or a tool with a known minimal system prompt.
  5. Treating all text as equal in token cost. Code and non-English content consume more tokens per character than English prose. Fix: estimate conservatively — use 2x your word count when working with code or mixed-language text.

Frequently Asked Questions

How many tokens is a typical ChatGPT conversation?

A short back-and-forth of ten exchanges runs roughly 1,000–3,000 tokens, well within any modern limit. A long research session with large pastes can exceed 50,000 tokens. I hit this regularly when pasting full articles for editing — the session grows faster than it looks.

Does a larger context window make the AI smarter?

Not directly. It means the model can consider more content at once, but reasoning quality depends on the model itself. A weaker model with a million-token window can still give shallow answers; a strong model with a smaller window often outperforms it on focused tasks.

Can the AI summarize itself when the context fills up?

Some apps do this automatically in the background, but the base models do not do it natively. If the context window fills, old messages get dropped silently — you will not receive a warning unless the app specifically shows one.

Is “context window” the same as “memory”?

No. Memory features (like ChatGPT’s persistent memory) store facts across sessions in a separate system, outside the context window. The context window is temporary — it resets with each new conversation.

Do tokens cost money on free plans?

On free tiers, token usage typically counts against a daily message or usage cap rather than direct billing. On paid API plans, you pay per 1,000 tokens consumed, so longer context windows can add up quickly on large tasks.

Conclusion

AI tokens and context windows explained simply: tokens measure the text, and the context window determines how much the AI can hold in view at once. Knowing this helps you pick the right tool, structure your prompts better, and understand why an AI sometimes seems to forget what you told it.

A good next step is trying the Custom GPT build guide — setting up your own GPT with a focused system prompt is one of the best ways to keep the context window free for the content that actually matters.

Build a Custom GPT in 5 Steps — No Coding Needed

Build a custom GPT without coding — open GPT Builder, describe your assistant, and share it in under 20 minutes with no technical background required.

Building a custom GPT felt like a developer task to me — something requiring API keys and Python experience. That assumption vanished the first time I opened GPT Builder inside ChatGPT and had a working assistant ready in eighteen minutes, no code at all. The key insight is that GPT Builder converts a plain-English description into a specialized AI assistant — your only job is explaining exactly what you need.

You do need a ChatGPT Plus, Team, or Enterprise subscription to build custom GPT assistants — the free tier can run existing ones, but not create new ones. Once you have access, the use cases are wide: a writing editor that follows your style guide, a customer FAQ bot loaded with your documentation, a study tutor that quizzes you on your own notes — all shareable with a single link.

Quick Answer

Open ChatGPT, click your profile icon → My GPTs → Create a GPT. Describe your assistant in the chat window, let Builder draft the instructions, then refine them in the Configure tab. Add a name, upload optional knowledge files, and click Save. Total setup time: under 20 minutes — no coding required.

What Is a Custom GPT?

A Custom GPT is a version of ChatGPT you preconfigure for a specific job. Instead of re-explaining your context at the start of every new chat, you set the role, tone, and rules once — and the assistant holds them every time you open it.

The GPTs you build can stay private, be shared by link, or be listed in the GPT Store for public discovery. Anyone with a free ChatGPT account can run a GPT you share — they don’t need Plus to use it, only to build their own.

A Custom GPT is essentially a saved system prompt with a name and avatar — nothing more technical than that.

How Do I Build a Custom GPT Step by Step?

Step 1: Open GPT Builder

Log into ChatGPT with a Plus, Team, or Enterprise account. Click your profile icon in the top-right corner, select My GPTs, and press Create a GPT. The screen has two tabs: Create (a guided chat) and Configure (manual field editing).

Step 2: Describe Your Assistant

Type what you want the GPT to do in the Create tab’s chat box. Builder asks follow-up questions and auto-fills the Configure fields as you answer. I typed “A writing assistant that rewrites any text in plain, friendly English and always asks who the audience is first.” Four exchanges later it had a complete draft — name, description, and starter instructions all generated.

Step 3: Review the Configure Fields

Switch to Configure to review and edit what Builder produced. Here is what each field controls:

Field What It Controls Required?
Name & Description Displayed in the GPT Store and link previews Yes
Instructions Core role, rules, and tone for every conversation Yes
Conversation Starters Sample prompts shown at the top of each new chat No
Knowledge Files the GPT can search at runtime (PDF, CSV, TXT) No
Capabilities Web Search, Image Generation, Code Interpreter toggles No

Step 4: Upload Knowledge Files (Optional)

If your GPT needs domain-specific material — a product catalog, a company FAQ, a class syllabus — upload those files under Knowledge. ChatGPT searches them at runtime. The limit is 20 files, up to 512 MB each.

Step 5: Save and Test

Click Save, pick a visibility level (Only me, Anyone with the link, or Public), and use the live preview panel on the right to send test prompts. Include off-topic and edge-case messages to confirm the GPT holds its rules under pressure.

Pro tip: Use the Create tab to generate a first draft of instructions fast, then switch to Configure for precise edits. Direct editing in Configure is quicker once you know exactly what you want to change.

The five steps cover the complete build cycle: open, describe, configure, optionally add knowledge files, then save and stress-test.

What Should I Write in the Instructions Field?

The Instructions field drives everything your GPT does. I treat it like a job description: role, audience, required behavior, and firm guardrails. A template that has worked well for me:

You are [Name], a [role] for [audience]. Always [required behavior]. Never [prohibited behavior]. When you don’t know something, say so clearly — don’t guess.

Keep instructions under 8,000 characters — the model deprioritizes rules buried at the end of very long prompts, so put your most important rule first. Bullet points stick better than prose for lists of rules. If you want account-wide tone preferences that apply to every regular chat, pair your GPT with ChatGPT’s account-level custom instructions — the two work independently and reinforce each other.

Troubleshooting tip: If your GPT keeps ignoring a specific rule, move that rule to the very first sentence of the Instructions field. The model weights the opening of the system prompt most heavily.

Instructions are the heartbeat of your GPT — a precise role and firm guardrails beat a long, vague prompt every time.

How Do I Share and Publish My Finished GPT?

At save time you choose a visibility level. “Only me” keeps it private for personal use. “Anyone with the link” generates a shareable URL — anyone with a free ChatGPT account can open and run it. “Public” submits your GPT to the GPT Store for discovery by all ChatGPT users.

To grab the link at any time, open My GPTs, click the three-dot menu next to your GPT’s name, and select Copy Link. That URL opens your assistant directly, with no extra navigation required.

Sharing takes one click — your audience only needs a free ChatGPT account to use what you built.

What Mistakes Should I Avoid?

  • Vague instructions. “Be helpful and friendly” is not a role. Name the specific task, the target user, and the expected output format. The more concrete the instructions, the more consistent the behavior.
  • Overloading knowledge files. Fifteen loosely related PDFs confuse document retrieval. Upload only files directly relevant to the GPT’s single job.
  • Skipping the test phase. Always send five to ten test prompts — including off-topic and adversarial messages — before sharing the GPT with anyone else.
  • Forgetting to update files. If your source material changes (a new price list, an updated policy), re-upload the file manually. The GPT does not auto-sync.
  • Confusing Custom GPTs with account custom instructions. Account-level custom instructions shape every regular chat you start. A Custom GPT is a separate, shareable assistant you open on demand — different tools with different scopes.

Most Custom GPT problems trace back to the Instructions field — write it like a precise job description and the common failure modes disappear.

Frequently Asked Questions

Do I need a paid ChatGPT plan to build a Custom GPT?

Yes — GPT Builder requires ChatGPT Plus ($20/month), Team, or Enterprise. Free accounts can run Custom GPTs shared by others but cannot create new ones. I upgraded to Plus specifically for this feature.

Can people without ChatGPT Plus use a GPT I share?

Yes. Anyone with a free ChatGPT account can open and use a GPT you share by link. Only the person building the GPT needs a paid plan — not the people you share it with.

How many Custom GPTs can I create?

OpenAI has not published a hard cap. I’ve built over a dozen on a single Plus account without hitting any limit. In practice, you are unlikely to reach a ceiling.

Can my Custom GPT search the web?

Yes, if you enable Web Search under Capabilities in the Configure tab. The GPT will pull live search results when the query needs current information — I keep this on for any assistant that covers fast-moving topics.

Are the files I upload used to train ChatGPT?

No — OpenAI uses uploaded files for retrieval within your GPT’s conversations only, not to train the base model. Review OpenAI’s privacy policy before uploading sensitive business documents.

These five questions cover what most people ask when they open GPT Builder for the first time.

Conclusion

Building a custom GPT is one of the fastest ways to make ChatGPT genuinely useful for a specific recurring task. My first build took eighteen minutes; every iteration since has been faster. Once you have one working, explore what else you can automate — summarizing long PDFs with ChatGPT pairs especially well with a document-review GPT tuned to your style. Pick one clear use case, build the assistant, and refine it as you use it.

NotebookLM vs ChatGPT for Research: What I Learned Using Both

NotebookLM vs ChatGPT for research: after testing both tools daily, I found a clear winner for source-based work — and a different answer for open exploration.

When I first started using AI tools for research, the same frustration kept coming up: the chatbot I reached for would either fabricate details from my source material or give me advice that completely ignored the document I’d uploaded. The two tools people compare most for notebooklm vs chatgpt research tasks — Google’s NotebookLM and OpenAI’s ChatGPT — are built for different jobs, and picking the wrong one wastes real time.

The core difference is this: NotebookLM anchors every answer to the documents you upload and shows you the exact citation, while ChatGPT draws from broad pre-trained knowledge and may blend that with your file’s content. That one distinction is the whole decision.

Quick Answer

NotebookLM is the better choice when you need verified, source-cited analysis of documents you already have. ChatGPT is better for open-ended research, brainstorming, and drafting. Both offer free plans. Use NotebookLM for precision; use ChatGPT for range. If you’re still mapping your topic, start with ChatGPT and move to NotebookLM once your sources are gathered.

What Is NotebookLM, and What Makes It Different From a Chatbot?

NotebookLM is a free AI research tool from Google that works exclusively with sources you supply. You upload PDFs, Google Docs, websites, YouTube links, or audio files, and it builds a private notebook from them. Every answer includes an in-line citation you can click to jump to the exact passage in your source.

How Source-Grounding Changes the Research Experience

I tested NotebookLM on a 50-page technical report and it quoted the precise paragraph without adding any outside context the document didn’t contain. For academic writing, legal analysis, or any work where a single wrong statistic matters, that constraint is a feature. The risk of a hallucinated fact is dramatically lower than with a general chatbot because the model simply can’t venture outside what you’ve given it.

NotebookLM’s strict source fidelity makes it the safer tool whenever you need every claim to trace back to a real document passage.

How Does ChatGPT Handle Research Tasks?

ChatGPT draws on a large pool of pre-trained knowledge and, on paid plans, can browse the web to supplement it. I reach for it early in a project — when I’m generating research questions, drafting an outline, or asking “what are the main debates in X field?” before I’ve gathered a single source.

What Happens When You Upload a File to ChatGPT?

Paid ChatGPT plans let you attach PDFs and documents for direct analysis. The key difference from NotebookLM: ChatGPT may blend your document’s content with its training data, producing answers that sound authoritative but mix cited and generated context. I caught it restating a figure from my report slightly off once because it interpolated from related background knowledge — exactly the scenario that makes NotebookLM safer for citation-critical work.

ChatGPT is strongest for open, exploratory research and writing where its broad knowledge base adds real value — not for strict, verifiable, document-only analysis.

How Do NotebookLM and ChatGPT Compare Side by Side?

Feature NotebookLM ChatGPT
Inline source citations Yes, per passage No (file analysis only)
Web browsing No Yes (paid plans)
Free plan Yes, fully featured Yes, with usage caps
File types accepted PDF, Docs, websites, audio, YouTube PDF, images, code files
Best use case Deep analysis of your own sources Open research and drafting

Pro tip: Use ChatGPT to identify which sources to find, then upload those sources to NotebookLM for citation-backed deep analysis. The two tools complement each other across different phases of the same project.

Neither tool dominates every research scenario — the right choice depends on whether you’re exploring broadly or drilling into a fixed set of sources you already have.

When Should You Choose NotebookLM Over ChatGPT?

Use NotebookLM when you have your documents in hand and need verifiable, cited answers — literature reviews, case studies, legal document review, or any work where a fabricated fact creates real damage. My go-to starting prompt there is: “Summarize the main argument of each source and note where they disagree.”

Use ChatGPT when you’re early in a project and need to think, draft, and explore. The guide on writing effective ChatGPT prompts covers the techniques that matter most for research-style queries. For real-time cited web searches, Perplexity AI is a third option worth keeping alongside both tools.

Troubleshooting tip: If NotebookLM returns “can’t find relevant information,” your PDF may be a scanned image without selectable text. Run it through a free OCR converter and re-upload the text-based version.

Matching the tool to the research phase — exploration vs. deep source analysis — saves more time than any prompt optimization trick.

Common Mistakes to Avoid

  1. Trusting ChatGPT for exact statistics from an uploaded file. It can blend document data with its training. Fix: use NotebookLM’s citation links to verify any figure before publishing it.
  2. Uploading scanned-image PDFs to NotebookLM. Image-only files don’t parse as text. Fix: use a PDF with selectable text, or run it through a free OCR tool first.
  3. Expecting NotebookLM to browse the web. It has no internet access at all. Fix: use ChatGPT with Browse enabled, or use Perplexity AI for live research.
  4. Assuming ChatGPT’s file-upload feature equals NotebookLM. Citation depth is shallower and hallucination risk is higher. Fix: for source-critical work, always use NotebookLM even when you already have ChatGPT open.
  5. Skipping NotebookLM’s Audio Overview. It turns your notebook into a podcast-style discussion of your sources — a fast way to absorb long documents during a commute.

Frequently Asked Questions

Is NotebookLM really free to use?
Yes. NotebookLM is free through Google, no credit card needed. The paid NotebookLM Plus tier adds higher limits, but the free plan supports up to 100 notebooks with 50 sources each. For most research projects, the free plan is more than enough — I’ve never hit the limit on a normal workweek.

Can ChatGPT replace NotebookLM for document research?
Not completely. ChatGPT reads files but lacks NotebookLM’s per-passage citation links and is more likely to mix document content with trained knowledge. When I write anything that gets published, I always verify figures in NotebookLM before including them.

Does NotebookLM support YouTube videos as sources?
Yes. Paste a YouTube URL and NotebookLM parses the transcript automatically. I use this for conference talks and long lectures — I can ask specific questions about a two-hour video in seconds instead of scrubbing through it.

Can I use both tools on the same research project?
Yes, and I recommend it. I use ChatGPT to outline research questions and identify key source titles, then upload those sources to NotebookLM for deep, citation-backed analysis. The workflow takes five minutes to set up and pays off on every long project.

Conclusion

For structured, source-backed research, NotebookLM is the clearer choice — it cites precisely, stays grounded, and is completely free. For open exploration and writing, ChatGPT handles the range. Use them together: start in ChatGPT, finish in NotebookLM. If you want to get more from ChatGPT in the meantime, the guide on summarizing PDFs with ChatGPT is a natural next step.

How to Summarize a PDF With ChatGPT in Minutes

Summarize a PDF with ChatGPT in minutes — upload the file, write a focused prompt, and get a plain-language overview with no extra software required.

Staring at a 60-page research paper you need to absorb in the next hour is exactly the kind of situation ChatGPT was built for. Whether it’s a legal contract, a technical report, or a dense academic study, the fastest path from an overwhelming document to a usable summary is a direct file upload to ChatGPT — no copy-paste, no third-party paraphrasing tools, no extra software.

I’ve run this workflow on quarterly financial reports, conference papers, and dense policy documents. It reliably saves me 20 to 30 minutes per document when I use a specific, targeted prompt instead of a generic one.

Quick Answer

To summarize a PDF with ChatGPT, open a new chat at chat.openai.com, click the paperclip icon, upload your file, and type “Summarize this document in plain language.” ChatGPT reads the file directly and returns a condensed overview in seconds. File uploads work on both the free and Plus plans.

A text-based PDF plus a targeted prompt is the full recipe — no paid plan required for most documents.

Which PDF Types Give the Best Results?

Not every PDF feeds cleanly into ChatGPT. The single biggest factor is whether your file contains a real text layer. If you can highlight and copy text normally inside your PDF reader, the file will work well.

PDF Type Works? Notes
Text-based (exported from Word or Google Docs) Yes — best results Native text is read precisely
Scanned documents with an OCR layer Usually Accuracy depends on scan quality
Image-only scanned PDFs (no OCR) No Invisible to the model; convert first
Password-protected PDFs No Remove the password before uploading
Slide decks saved as PDF Yes ChatGPT reads each slide’s text

Text-based PDFs give precise summaries; image-only files must be converted with an OCR tool before ChatGPT can read them.

How Do I Summarize a PDF With ChatGPT?

Step 1: Open a New Chat

Go to chat.openai.com and sign in. Click New chat in the left sidebar. Starting a fresh conversation prevents earlier context from influencing your summary.

Step 2: Upload the PDF

Click the paperclip icon (or the plus button on mobile) next to the message box. Select your file. The PDF appears as a thumbnail once it finishes uploading.

Pro tip: Files over 25 MB may slow the upload. Compress large PDFs first with a free tool like Smallpdf or PDF24 before attaching them.

Step 3: Write a Focused Prompt

Type your instruction in the same message box before sending. These prompts consistently outperform a bare “summarize this”:

  • “Summarize the main findings in five bullet points.”
  • “What are the key obligations and deadlines in this contract?”
  • “Explain this report to someone with no background in finance.”

I tested a 40-page market research report with a generic “summarize this” prompt versus “summarize the competitive landscape section in five bullets for a business owner.” The targeted version returned five directly actionable insights; the generic one produced a dense, broad paragraph that required further reading to use.

Step 4: Follow Up to Drill Deeper

ChatGPT holds the document in context for the full conversation. After your initial overview, ask sharper follow-up questions:

  • “List every date or deadline mentioned.”
  • “Explain section 3 in simpler terms.”
  • “Quote the exact passage where the refund policy is described.”

That last prompt is especially useful for fact-checking — it forces ChatGPT to anchor its answer in the source text rather than paraphrase loosely.

Troubleshooting tip: If ChatGPT says it can’t read your file, the PDF is likely image-only. Open the file in Google Drive, let Drive convert it to a Google Doc (which triggers automatic OCR), then download the resulting text version and re-upload it to ChatGPT.

Step 5: Save Your Summary

ChatGPT does not auto-save your output. Copy the summary into your notes app or a Google Doc before closing the tab. The free plan’s conversation history may not persist across sessions.

A targeted prompt in Step 3 is the single biggest lever for improving summary quality — it matters more than which ChatGPT plan you use.

How Can I Get More Focused Summaries?

What Prompt Adjustments Make the Biggest Difference?

Three changes consistently sharpen ChatGPT’s output when summarizing PDFs:

  • Narrow the scope: “Summarize only the methodology section” beats “summarize this document.”
  • Set a format: “Return a table with columns for Topic and Key Finding.”
  • Define the audience: “Explain this to a non-technical hiring manager.”

For very long documents, ask ChatGPT to navigate by section heading or chapter name rather than page number. It handles headings more reliably than absolute page counts, which it sometimes miscounts in dense, multi-column layouts.

Specificity in the prompt consistently produces narrower, more useful summaries than open-ended requests — this holds true whether the document is 5 pages or 500.

Is ChatGPT Reliable Enough for Important Documents?

ChatGPT handles most documents accurately, but it can miss details buried in footnotes, dense tables, or heavily formatted text. For any figure, date, or clause that matters, verify it against the original. I treat ChatGPT’s summaries as an orientation tool — a starting point, not a final answer — especially for legal or financial documents.

If you need a second opinion on a sensitive file, see how Claude AI handles document uploads on its free plan as an alternative worth testing.

Use the summary to orient yourself quickly, then go back to read any critical section in full before acting on it.

What Common Mistakes Should I Avoid?

  1. Uploading a password-locked PDF. ChatGPT cannot open it. Remove the password first with your PDF reader or a free tool like iLovePDF, then re-upload.
  2. Using an image-only scanned PDF. The file looks normal to you but contains no text layer. Convert it via Google Drive or Adobe Acrobat before uploading.
  3. Sending a vague prompt. “Summarize this” returns a generic overview. Specify the section, format, or audience to get output you can act on immediately.
  4. Trusting numbers without spot-checking. ChatGPT occasionally misses figures buried in tables or footnotes. Always verify key data points against the original document.
  5. Closing the tab without copying your summary. The free plan does not reliably save session history between logins. Copy before you navigate away.

Frequently Asked Questions

Does ChatGPT’s free plan support PDF uploads?

Yes. File uploads work on the free tier with a daily limit on the number of files. If you hit the cap, ChatGPT will prompt you to try again later or upgrade to Plus. For occasional use, the free plan handles most needs without any cost.

Is my uploaded PDF kept private?

OpenAI does not use uploaded files to train its models by default. For sensitive documents — contracts, medical records, financial statements — I recommend removing personal identifiers before uploading and reviewing OpenAI’s current privacy policy for the full terms.

Can I upload multiple PDFs at once?

Yes. Attach several files to one message and ask ChatGPT to compare, contrast, or find common themes across them. I’ve summarized three competing vendor proposals in a single chat this way and received a clean side-by-side breakdown without any manual formatting.

What should I do if the summary contains errors?

Ask ChatGPT to quote the source text directly: “Quote the exact passage where this claim appears.” If no clear quote exists, the detail was likely hallucinated or missed. Follow up with a targeted question about that specific section to dig out the accurate information.

Conclusion

Summarizing a PDF with ChatGPT is one of the most practical uses of the tool today — and it works on the free plan. Pair a clean, text-based file with a specific prompt and you can turn a dense document into a usable set of notes in under five minutes.

If you work inside Google’s ecosystem, see how Google Gemini summarizes documents directly inside Google Docs. Windows users with Microsoft 365 should check what Microsoft Copilot can do inside Word — no file upload needed.

Start with a short PDF you already need to read. Upload it, try one of the targeted prompts above, and see how much time you save.

ChatGPT Cuts Off Mid-Response: Why It Happens and How I Get the Full Answer

ChatGPT cuts off mid-response when it hits the output token limit. Here is why it stops and the quick steps I use to recover the full answer without paying.

I was halfway through a detailed answer from ChatGPT when the text just stopped. No error, no warning, no spinning dots. The reason ChatGPT cuts off mid-response is almost never a broken account or a bug on your end. It is the output token limit doing exactly what it was designed to do, and you can recover the rest of the answer in seconds.

Over months of daily use I have hit this cut-off hundreds of times while drafting long articles and analyzing reports. The cause is usually one of three things: an output token cap, a network hiccup, or server throttling at peak hours. Once you know which one hit you, the fix takes under a minute and costs nothing.

Quick Answer

Type Continue in the same chat box and press send. ChatGPT reads the conversation history and resumes right where it stopped. If it starts fresh instead, ask it to finish from the exact point it cut off, or request a shorter format such as bullet points or a 300-word summary to stay under the cap.

Why Does ChatGPT Cut Off Mid-Response?

There are three common causes, and each one leaves a different fingerprint on where and how the text stops.

Token Output Limits

Every response is capped at a maximum number of tokens, roughly 4,096 for GPT-4o on the free tier. One token is about three-quarters of a word, so a dense technical explanation or a long draft can hit that ceiling mid-sentence. This is the most common cause I see, and it is by design.

Network Timeouts

A slow or unstable connection can interrupt the streaming of a reply before it finishes. When my answer dies at a random spot mid-word rather than at a natural stopping point, I know it was the network, not the token limit.

Server-Side Throttling

During peak hours, usually weekday afternoons in North American time zones, OpenAI handles a flood of simultaneous requests. Free users can get shorter replies as the system balances load. Paid tiers get priority access that mostly avoids this.

The spot where ChatGPT stops tells you the cause: mid-sentence means the token cap, mid-word means the network, and shorter-than-usual replies point to peak-hour throttling.

How Do I Get the Full Answer When ChatGPT Cuts Off?

Work through these steps in order. The first one resolves it for me the vast majority of the time.

Ask It to Continue

Type Continue in the chat box and send. ChatGPT picks up from where it stopped, usually in under five seconds on a stable connection. You can repeat this as many times as you need, because the context window still holds the previous output. If Continue starts a new response, I use this instead: “Please finish the previous response, starting from where you stopped.”

Request a Shorter Format

Before re-sending a long prompt, add a format instruction such as “Answer in bullet points, each under 25 words” or “Give me a 300-word summary.” This keeps the whole answer inside the token limit, so it never cuts off in the first place. A clear, specific prompt also helps; see my guide to writing ChatGPT prompts like a pro for the formatting tricks I rely on.

Split Your Prompt Into Parts

For genuinely long jobs, like a full essay or a large document analysis, I break the task into sections. I ask for the introduction first, then the body, then the conclusion. Each section stays well under the cap, and the focused output is cleaner at every step.

Refresh and Retry at Off-Peak Hours

If the cut-off lands at the same spot no matter how short the prompt is, server load is the likely cause. Refresh the page, wait 30 seconds, and try again. Early morning or late evening gives me far fewer interruptions than mid-afternoon. Before retrying, I check the OpenAI status page; if there is a listed incident, waiting beats troubleshooting.

Compare Plans and Upgrade Only if Needed

If you hit the output limit constantly, the table below shows the practical differences across tiers.

Plan Model Access Approx. Max Output Peak-Hour Priority
Free GPT-4o (rate-limited) ~4,096 tokens Low
Plus ($20/mo) GPT-4o, o1 ~16,000 tokens High
Team ($25/user/mo) GPT-4o, o1 ~16,000 tokens High
API (pay-as-you-go) All models Up to 128k tokens Configurable

For most free users, the Continue command erases the problem entirely. If you keep hitting the limit on Plus or Team, the OpenAI API with a high max_tokens value gives you full control over output length.

Start with Continue, fall back to a shorter format, and only consider a paid tier if you truly hit the ceiling every day.

What Mistakes Should I Avoid When ChatGPT Cuts Off?

These are the missteps that cost me the most time before I understood what was really happening, along with the fix I now use for each one.

  1. Re-sending the full prompt. This starts a brand-new response instead of a continuation. Fix: type Continue in the same chat thread to resume.
  2. Opening a new chat window. A new conversation loses all prior context. Fix: stay in the same thread and use the continue command.
  3. Assuming it is a bug. Token-limit cut-offs are expected behavior. Fix: treat them as a format problem, not an error, and skip the unnecessary troubleshooting.
  4. Pasting huge chunks of text in one prompt. Large inputs eat tokens that would otherwise go toward the answer. Fix: break big pastes into smaller pieces.
  5. Ignoring the status page. If three retries all stop at the same point, the problem is on OpenAI’s side. Fix: check the status page before spending more time on workarounds.

Almost every wasted minute here comes from starting over instead of resuming inside the same thread.

Frequently Asked Questions

Why does ChatGPT stop mid-sentence?
It hit its token output limit. For example, when I asked for a 2,000-word breakdown of a contract, it stopped cleanly mid-sentence around the token ceiling, and a single Continue finished the rest.

Does typing Continue always work?
It works reliably when the cause is a token limit. For instance, after I refreshed the page during one long answer, Continue started a new reply, so I had to ask it to resume from the last line instead.

Will upgrading to ChatGPT Plus stop the cut-offs?
It greatly reduces them but does not remove them entirely. When I moved to Plus, my long research replies stopped cutting off during weekday afternoons thanks to the higher ceiling and priority access.

Is this the same as ChatGPT not loading at all?
No. A cut-off means the answer started and stopped, while a failure to load is a connectivity issue. The day my page would not open at all, the steps in how to fix ChatGPT when it stops working got me back in.

Can I fix this on the ChatGPT mobile app?
Yes. Type Continue in the chat box on iOS or Android and it resumes just like desktop. I have finished long answers from my phone on the train this way more times than I can count.

Does switching to a different AI chatbot help?
Some models have higher default output limits, but the same token concept applies everywhere. When I needed longer single replies, I compared options in ChatGPT vs Gemini vs Claude before deciding.

What Should I Do First When the Text Stops?

A ChatGPT cut-off is rarely a sign that anything is broken. It is the output token limit working as intended, and a quick Continue or a shorter-format request clears it in seconds for nearly every case I run into.

Next time the text stops, type Continue before you do anything else, then come back and bookmark this page so the fix is one click away.

Resume in the same thread first, reshape the format second, and reach for a paid tier only when the ceiling truly blocks your daily work.

Free AI Image Generators Compared: DALL-E 3 vs Firefly vs Canva AI

I compare the best free AI image generator tiers — DALL-E 3, Firefly, and Canva AI — so you pick the right one and write prompts that land usable images fast.

I have generated hundreds of images across the free tiers of DALL-E 3, Adobe Firefly, and Canva AI, and the gap between them comes down to one thing: matching the tool to the job. Pick the wrong free AI image generator for your task and you waste limited credits on results you cannot use.

Each tool hands you a different free allowance, a different output style, and a different licensing story. Below I break down exactly what each free tier gives you, where each one wins, and the prompt structure I rely on to land a usable image on the first or second try.

Quick Answer

DALL-E 3 (via the ChatGPT free tier) gives roughly 2 images per day with strong photorealism. Adobe Firefly offers 25 free generative credits monthly with the safest commercial licensing. Canva AI Text to Image provides 50 lifetime free uses and the easiest interface. For most non-commercial tasks, I start with Canva.

What Does Each Free Tier Actually Give You?

Before choosing, I look at three numbers: how many free images you get, what style each tool favors, and what each one is genuinely best at. Here is how the three compare side by side.

Tool Free Credits Image Style Best For
DALL-E 3 (via ChatGPT) ~2 images/day Photorealistic, artistic Varied or complex prompts
Adobe Firefly 25 credits/month Clean, polished Commercial-safe work assets
Canva AI Text to Image 50 lifetime uses Illustrative, design-ready Social media and slides

Each tool occupies a clear lane, so the right pick depends entirely on your task and licensing needs.

How Good Is DALL-E 3 on the ChatGPT Free Plan?

OpenAI built DALL-E 3 directly into ChatGPT, so you reach it through the same account you may already use. The free plan allows roughly 2 to 3 image generations per day before the cap hits. In my experience a generation takes 10 to 20 seconds; during peak US evening hours I have waited closer to 45 seconds.

What sets DALL-E 3 apart is how well it follows nuanced, detailed prompts. Describe a specific mood, an unusual combination, or a stylistic reference and it usually delivers. You can also refine the same image conversationally — I typed “make the sky more dramatic” once and it adjusted without starting from scratch.

I get faster refinements when I pair generation with solid prompt habits from ChatGPT Custom Instructions, which set my preferred tone and context upfront.

DALL-E 3 is the free tool I reach for when a prompt is complex or unusual.

Is Adobe Firefly Safe for Commercial Work?

Adobe Firefly is trained on Adobe Stock images and public-domain content, which makes its output commercially safe — a critical distinction for business or client work. A free Adobe account gives 25 generative credits per month, and each standard image costs 1 credit.

The browser-based interface at firefly.adobe.com includes style sliders for content type, color, lighting, and composition. I find that far friendlier for beginners than staring at a blank prompt box. When I produced a header for a paid client deck, Firefly was the only free tier I trusted for licensing.

Firefly is my pick whenever the image has to clear commercial use.

How Firefly Handles Credits

Because credits reset monthly rather than carrying over, I treat them as a budget. I prototype elsewhere and save Firefly for the final, polished export I actually intend to ship.

When Should I Use Canva AI Text to Image?

If you already use Canva for social media or presentations, the built-in Text to Image tool is the fastest way to generate an image and drop it straight into your design. The free plan includes 50 lifetime generations — not monthly — so I use them intentionally.

Canva AI performs best on illustrative or graphic-design-style prompts rather than photorealistic scenes, which suits its core audience perfectly. I lean on it for slide graphics and social posts where the image lives inside a layout anyway.

Canva is the quickest path from prompt to finished design.

How Do I Write Prompts That Get Usable Results?

Vague prompts produce vague images. Every effective prompt I write has three parts:

  1. Subject — who or what is in the image (“a golden retriever puppy”)
  2. Context — where and how (“sitting on a wooden porch at sunrise”)
  3. Style cue — the visual look (“photo-realistic, shallow depth of field, warm tones”)

Put together: “A golden retriever puppy sitting on a wooden porch at sunrise, photo-realistic, shallow depth of field, warm golden tones.” That single prompt reliably returns usable images across all three tools.

If results look distorted or overcrowded, I strip the prompt back to one subject and one style cue, generate once, then add context in a follow-up. For broader options, I also keep this roundup of free AI image tools handy when a project needs a fourth or fifth alternative.

Subject, context, and style cue together turn a guess into a repeatable result.

What Are the Most Common Mistakes to Avoid?

  • Using one-word prompts. “Cat” returns a generic cat. The fix: add setting, mood, and style so the output is usable in a design.
  • Forgetting to set the aspect ratio first. The default is usually square (1:1). The fix: select 16:9 before generating a YouTube thumbnail or LinkedIn banner, because changing it afterward means starting over.
  • Expecting readable text inside the image. Generators consistently fail at legible words. The fix: generate the graphic, then add your text overlay in Canva or another design app.
  • Burning Firefly credits on early drafts. The fix: test your prompt in Canva or ChatGPT first, then switch to Firefly only for the commercial-safe final output.
  • Assuming all free outputs are copyright-free. The fix: review current platform terms — Firefly is designed for commercial use, while DALL-E 3 outputs follow OpenAI’s usage policies.

Avoid these five traps and almost every generation comes back usable on the first or second try.

Frequently Asked Questions

Do I need to pay to use DALL-E 3?
No. DALL-E 3 is available on the ChatGPT free plan with a limited number of daily generations. When I hit the cap on a busy afternoon, I simply waited until the next day rather than upgrading; ChatGPT Plus subscribers get priority access and higher limits.

Can I use Adobe Firefly images for my business?
Yes. Adobe trained Firefly on licensed content, so outputs are cleared for commercial use. I used a Firefly image in a paid client presentation and reviewed Adobe’s current Firefly terms first to confirm there were no industry-specific restrictions.

What happens when I run out of Canva AI credits?
Canva’s free plan includes 50 lifetime Text to Image generations, and further use requires Canva Pro. I burned through mine quickly by regenerating drafts, so now I finalize the prompt before clicking generate.

Which tool produces the most realistic photos?
DALL-E 3 generally leads for photorealism on complex or unusual scenes. When I needed a moody dusk cityscape, DALL-E 3 nailed it while Firefly stayed stronger for studio-style product and portrait shots.

Are there other free AI image generators worth trying?
Yes. Microsoft Designer, powered by DALL-E, is another solid free option. When I want source-backed comparisons of AI tools, I run the query through Perplexity AI because it surfaces cited reviews quickly.

Does the AI store the images I generate?
Policies differ by platform, and both OpenAI and Adobe offer data-usage controls in account settings. Before generating anything proprietary for work, I checked each platform’s privacy settings to keep the content out of training data.

Conclusion

Each tool has a clear lane: Canva for fast in-design graphics, Firefly for commercial-safe assets, and DALL-E 3 for the most flexible read of a complex prompt. Start with the subject-context-style structure and you will land a usable image quickly.

Pick one free tier today, run the three-part prompt, and see which output fits your next project. For more on what AI already does in your everyday apps, see what Google Gemini can do inside Gmail and Google Docs.

Perplexity AI: Get Cited, Real-Time Answers for Any Research Question

Perplexity AI cites every source it pulls in real time. I show how to search, pick Focus modes, build threads, and keep research organized with Collections.

Perplexity AI is an AI answer engine that searches the live web and shows you exactly which sources it used, with numbered citations on every answer. I switched to it after wasting an afternoon fact-checking a confident ChatGPT reply that turned out to be invented. Once you can see the receipts behind an answer, you stop trusting AI blindly and start verifying in seconds.

Unlike a traditional search engine that just lists links, Perplexity synthesizes those sources into a direct answer and then hands you the citations. It handles research questions, current events, comparisons, and technical topics, and the free tier covers most everyday use without a login.

Quick Answer

Perplexity AI is a free AI search tool that fetches real-time web results and cites every source it uses. Visit perplexity.ai, type a specific question, and check the numbered citations beside the answer. For deeper work, use follow-up questions to dig in and Collections to save and organize what you find.

How Does Perplexity AI Compare to ChatGPT and Google?

Before diving in, it helps to see where Perplexity fits next to tools you already use. Three or more options here are genuinely comparable, so I lined them up on the features that actually change your workflow.

Tool Real-time web Cites sources Conversational Free tier
Perplexity AI Yes Yes Yes Yes
ChatGPT (GPT-4o) Limited Rarely Yes Yes
Google Gemini Yes Partial Yes Yes
Google Search Yes N/A No Yes

Perplexity’s edge is combining real-time search and transparent citations in a conversational format. ChatGPT has added browsing, but its source attribution is inconsistent; if you want a fuller breakdown, my comparison of ChatGPT, Gemini, and Claude covers where each one wins. Google Search still just gives you links to read yourself.

Perplexity is the only one of these that pairs live results with visible citations by default.

How Do I Search on Perplexity AI?

Searching well comes down to three things: pick the right Focus mode, ask a full question, and read the citations instead of trusting the summary.

Open Perplexity and Choose a Focus Mode

Go to perplexity.ai. No account is required for most searches, but a free sign-up unlocks saved threads and Collections. Before typing, check the Focus selector — it defaults to Web. Other modes include Academic (peer-reviewed papers only), YouTube, and Reddit. Picking the right mode sharpens results immediately.

Write a Specific, Research-Style Query

Perplexity handles full questions better than single keywords. Instead of searching VPN, I ask “What’s the difference between a VPN and a proxy for privacy?” The more context you give, the more targeted the answer. Aim for a complete sentence that describes exactly what you need to know.

Read the Answer and Check the Citations

The answer appears in the center with inline citation numbers, and the sources panel lists the exact URLs — click any number to open that page. When a key claim looks off, I can confirm it in under 10 seconds, far faster than hunting for corroboration after an uncited answer.

Pro tip: Hover over a citation number before clicking. Perplexity shows a source preview snippet so you can judge credibility without leaving the page.

If results look stale: On a fast-moving topic, confirm Focus is set to Web and add the current year to your query — for example, “best free AI tools 2026.” That pushes Perplexity toward more recently indexed pages.

Choose a Focus mode, ask in full sentences, and verify through the citations rather than the summary.

How Do Follow-Up Questions and Threads Work?

Every Perplexity answer includes a follow-up field at the bottom that keeps full context from the previous exchange, so I can ask “Which of those options works offline?” without re-explaining the topic. A connected sequence of prompts is called a Thread, and signed-in users can save and revisit threads from the Library dashboard.

Ask one follow-up at a time — multi-part questions often get answers that cover only part of what you asked.

Stay in a single thread so each follow-up builds on the context already established.

How Do I Organize Research With Collections?

A Collection is a folder that groups related Threads. When I was comparing cloud storage options, I made a Collection called “Cloud Storage Research” and saved every related thread into it. Collections support collaborators too, which makes Perplexity practical for small team projects.

To create one: click Library in the left sidebar, choose New Collection, name it, then save threads into it from the Thread options menu.

Group threads into Collections so a multi-session research project stays in one place.

What Are the Most Common Perplexity AI Mistakes?

  • Treating citations as pre-verified truth. Perplexity points to real sources, but those sources can still be wrong or outdated. Check publication dates on anything that drives a real decision — my guide to fact-checking AI answers walks through the full workflow.
  • Using vague queries. “Tell me about AI” gets a surface overview. “How does retrieval-augmented generation differ from fine-tuning for enterprise use?” gets a precise, cited answer. The fix: write the full question.
  • Ignoring Focus modes. Leaving Focus on Web for academic topics mixes blog posts with peer-reviewed studies. Switch to Academic when source quality matters most.
  • Starting a new search for every related question. That loses context each time. Stay in the same thread so follow-ups build on what came before.
  • Assuming the free tier is too limited. Standard free searches handle most everyday research with no daily cap. Rate limits apply only to Pro searches, so reserve those for complex multi-step queries.

Avoid these five traps and Perplexity goes from a faster search box to a research tool you can actually trust.

Frequently Asked Questions

Is Perplexity AI free to use?
Yes. The free tier supports unlimited standard web searches with citations and needs no account to start. For example, I ran a week of daily research queries without signing in and never hit a wall. The paid Pro plan ($20/month) adds more Pro searches, advanced models, and file uploads.

Does Perplexity AI store my searches?
Only if you’re signed in — then threads are saved to your Library; searching logged out keeps the session off your profile. When I researched a sensitive medical question, I stayed logged out so nothing was tied to my account. For a useful comparison, see how ChatGPT handles stored data.

Can I use Perplexity AI for academic research?
Yes, and Academic Focus limits results to peer-reviewed papers from sources like PubMed, arXiv, and Semantic Scholar. When I needed citations for a paper, it surfaced three studies I’d missed on Google. Always read the original before citing it — the summary is a starting point, not a substitute.

How accurate is Perplexity AI compared to ChatGPT?
Both make mistakes, but Perplexity’s cited format makes errors far easier to catch. I once spotted a bad citation in seconds because the linked page didn’t support the claim. For tasks without live search, pairing strong prompt techniques with ChatGPT still works well.

Is there a Perplexity AI mobile app?
Yes — free iOS and Android apps are available. For example, I start a thread on my phone during a commute and finish it on my laptop, since threads and Collections sync automatically across devices.

Conclusion

Perplexity AI removes the biggest frustration with AI answers: you always know where the information came from. Ask specific question-style queries, lean on Focus modes, and build threads instead of starting over each time.

Run your next research question through perplexity.ai and compare the cited result to a plain search — the difference is hard to ignore.

ChatGPT Custom Instructions: Set Them Once, Get Better Answers Every Time

ChatGPT Custom Instructions personalize tone, context, and format once so every new chat starts smarter. Here is how I set mine up in three minutes flat.

Every time I open a fresh ChatGPT conversation, the AI starts blind — it has no idea what I do for work, the tone I prefer, or whether I want tight bullet points or a deeper walkthrough. I used to burn three or four messages just rebuilding that context before getting a single useful answer. ChatGPT Custom Instructions kill that warm-up tax permanently by storing your context once and applying it to every new chat.

Custom Instructions are a built-in feature on both free and paid ChatGPT accounts. You define two things — background about yourself and how you want ChatGPT to respond — and every new conversation inherits those preferences automatically. I set mine up in under three minutes, and my answers have started sharper ever since.

Quick Answer

Open ChatGPT, click your profile icon, then go to Settings, Personalization, and Custom Instructions. Fill the first field with your role and context, and the second with your preferred response style. Toggle Enable for new chats on, save, and every conversation you start after that uses your preferences automatically.

What Are ChatGPT Custom Instructions?

Custom Instructions are two persistent text fields stored with your OpenAI account:

  • “What would you like ChatGPT to know about you?” — your profession, skill level, location, or ongoing projects.
  • “How would you like ChatGPT to respond?” — tone (formal or casual), length (concise or detailed), format (bullets or prose), or any special preferences.

These act as a silent system prompt prepended to every new conversation. ChatGPT reads them before it reads anything I type, so I never repeat myself across sessions. When I added “freelance tech writer, plain-English voice” to mine, the AI stopped padding answers with disclaimers I never wanted.

Custom Instructions are a standing brief ChatGPT consults before your first word in every new chat.

Who Can Use Custom Instructions?

Custom Instructions are available on free ChatGPT accounts and ChatGPT Plus alike, so you do not need to pay to use them. The one exception is Temporary Chat mode, which is built for private sessions with no saved context and silences your instructions by design. The first time my answers felt oddly generic, I realized I had opened a Temporary Chat without noticing.

Every account tier supports Custom Instructions except Temporary Chat, which bypasses them on purpose.

How Do I Set Up Custom Instructions?

The path differs slightly between desktop and mobile, but both take a couple of minutes. I walk through each below.

On Desktop (Browser)

  1. Go to chatgpt.com and sign in.
  2. Click your profile icon in the bottom-left corner.
  3. Select Settings.
  4. Click Personalization in the left sidebar.
  5. Click Custom Instructions.
  6. Fill in both text fields (see the examples in the next section).
  7. Toggle Enable for new chats on, then click Save.

One thing tripped me up early: changes apply only to conversations started after you save, not to any chat already open. I now open a fresh tab right after saving to confirm the instructions took.

On the ChatGPT Mobile App (iOS and Android)

  1. Tap the three horizontal lines in the top-left corner.
  2. Tap your profile name at the bottom of the menu.
  3. Tap Custom Instructions.
  4. Fill in both fields and tap Save.

Setup lives under Personalization on desktop and your profile name on mobile, and only new chats inherit the changes.

What Should I Write in Each Field?

This is where most people stall. The key is specificity — vague instructions produce vague improvements. Below are examples I have tested across common roles so you can adapt one to your own work.

Use Case “Know About You” Field “How to Respond” Field
Developer Senior backend engineer, Python/Django, building B2B SaaS Skip basic explanations, show code first, explain trade-offs briefly
Student Undergrad biology student studying for exams Use simple language, add analogies, end with a 2-sentence summary
Non-technical professional Marketing manager, no coding background Avoid jargon, keep responses under 200 words unless I ask for more
Writer Freelance copywriter, B2B tech niche Match professional tone, prefer active voice, flag weak word choices
Parent / Educator Homeschooling parent, kids ages 8 and 11 Explain as if to a curious 10-year-old, use real-world examples

If your responses still feel generic after saving, confirm the Enable for new chats toggle is on and that you have opened a new conversation. Instructions will not retroactively apply to a chat that was already open when you saved. To go further on phrasing, my guide to writing ChatGPT prompts like a pro pairs well with a solid instruction set.

Specific role details and 2-3 clear response rules beat long, hedged paragraphs every time.

What Are the Common Mistakes to Avoid?

I made most of these before they became habits to skip. Here is what trips people up and the fix for each.

  1. Writing too much. Each field holds up to 1,500 characters, but cramming in every detail makes ChatGPT deprioritize what matters. Fix: stick to your 2-3 most important facts and 2-3 core response preferences.
  2. Contradicting yourself. Asking for “concise responses” in one field and “always include detailed step-by-step breakdowns” in the other creates unpredictable output. Fix: decide your priority and stay consistent.
  3. Never revisiting them. Your role, projects, and goals change, and stale instructions are almost as unhelpful as none. Fix: set a reminder to review them every few months.
  4. Storing sensitive data. Passwords, financial details, and private health information get sent on every prompt and stored on OpenAI’s servers. Fix: keep that data out of these fields entirely.
  5. Forgetting Temporary Chat silences them. A session opened in Temporary Chat bypasses your instructions by design. Fix: check the top of the chat window when answers feel impersonal.

To see the bigger picture of what ChatGPT stores and how it builds a profile from your chats, read my guide on what ChatGPT remembers and how to take back control.

Most failures come from overloading the fields or contradicting yourself, and each has a one-line fix.

Frequently Asked Questions

Do Custom Instructions work with GPT-4o and other models?
Yes, they apply regardless of which model you select. When I switch between GPT-3.5 and GPT-4o mid-project, the same instructions feed in as a system prompt before my first message every time.

Can people see my Custom Instructions if I share a chat link?
No, shared links display only the visible messages in that specific chat. I have shared dozens of conversation links with editors, and my instructions never appeared in any of them.

How are Custom Instructions different from ChatGPT’s memory feature?
Custom Instructions are static text you write yourself, while memory automatically saves facts ChatGPT picks up as you talk. I keep both on — my instructions set the voice, and memory remembers details like the project names I mention.

Will Custom Instructions affect image generation requests?
Partially — tone and style preferences will not change image output, but role context can shape how ChatGPT interprets an ambiguous prompt. When I ask for a “diagram,” my “tech writer” context nudges it toward clean, labeled visuals.

Can I have different Custom Instructions for different projects?
Not natively, since there is one global set per account. As a Plus user I use Projects, which let me override the global set inside a workspace — I keep a separate, code-first instruction set in my development project.

What happens if I delete my Custom Instructions?
ChatGPT reverts to default behavior as if none existed, and your open conversations are unaffected. I cleared mine once to troubleshoot a weird tone issue, then pasted them back in seconds.

Conclusion

Custom Instructions offer one of the best setup-to-payoff ratios in any AI tool — a few minutes of configuration that improves every conversation afterward. Fill both fields thoughtfully, save, and open a new chat to feel the difference right away. Still weighing your options? My ChatGPT vs Gemini vs Claude comparison breaks down which free AI fits your workflow next.

ChatGPT Memory: How to See, Delete, and Control What It Saves About You

ChatGPT memory quietly builds a profile of you. I show how to see every saved entry, delete what is stale, and stay private with Temporary Chat in under two minutes.

ChatGPT memory sounds convenient, and honestly it is. Once it is on, ChatGPT quietly notes your name, your job, your writing-style preferences, and personal context you mention in passing. The catch is that most people never see what has piled up. If you have used ChatGPT regularly for a few months, it almost certainly knows more about you than you would guess.

The first time I opened my own memory list, I found a half-marathon I had trained for the year before still shaping every health answer it gave me. Checking the full list takes about 90 seconds, and you can delete anything outdated or too personal. Below I show exactly where the list lives, how to remove items, and when to use Temporary Chat so nothing is saved at all.

Quick Answer

Open ChatGPT, go to Settings → Personalization → Manage memories to see everything saved about you. Delete single items with the trash icon, or toggle Memory off completely. For a private, one-off chat, click the pencil icon at the top of the sidebar and choose Temporary Chat so nothing from that session is kept.

In short, your full ChatGPT memory list is one settings panel away, and you can prune or disable it in seconds.

What Does ChatGPT Memory Actually Save?

When you use ChatGPT on the free or Plus plan, it can store facts you mention, then surface them automatically in later conversations. OpenAI introduced the memory feature for Plus subscribers in early 2024 and later extended it to free-tier accounts.

You never actively approve what gets saved. ChatGPT decides what is worth keeping, so entries accumulate silently. One casual line like “I am training for a half-marathon” can sit in your list for months, quietly steering every health-related reply you get.

The Categories It Tends to Remember

  • Personal details (name, job title, city or country)
  • Communication and formatting preferences
  • Ongoing projects or long-term goals
  • Details about family members or colleagues you mentioned
  • Technical context (programming languages, tools, platforms)

In short, anything you say in passing can become a permanent fact unless you remove it.

How Do I Check What ChatGPT Has Stored?

  1. Open ChatGPT at chatgpt.com and sign in.
  2. Click your profile icon in the bottom-left corner.
  3. Select Settings.
  4. Go to Personalization → Manage memories.
  5. Scroll the list; each entry shows the exact text ChatGPT saved.

The same list lives in the mobile app: tap your profile icon (top right) → Settings → Personalization → Manage memories. I run this check on the first of every month, and it takes me under two minutes.

The full list is one settings panel away on both web and mobile.

How Do I Delete or Control ChatGPT Memories?

Delete a Single Memory

  1. Open Manage memories using the steps above.
  2. Find the entry you want gone.
  3. Click the trash icon next to it and confirm.

It is removed instantly. There is no undo and no recycle bin; the entry is simply gone and stops influencing replies.

Clear Every Memory at Once

  1. Open Manage memories.
  2. Click Clear all memories at the top of the panel and confirm.

The whole list is wiped. If memory stays enabled, ChatGPT starts a fresh list from your next conversation.

Turn Memory Off Entirely

  1. Go to Settings → Personalization.
  2. Toggle Memory off.

ChatGPT stops saving and ignores the existing list, but your stored entries are preserved if you switch it back on later. If you would rather guide ChatGPT deliberately instead of letting it guess, set up ChatGPT Custom Instructions, and sharpen your wording with these prompt techniques for smarter responses.

Use Temporary Chat for Private Sessions

  1. In the sidebar, click the pencil (compose) icon at the top.
  2. Select Temporary Chat.
  3. Chat normally; nothing is saved to memory or history.

I keep Temporary Chat for sensitive questions, such as finances or health, where I want the help without leaving a footprint.

You have four levers here: delete one, clear all, switch memory off, or go temporary.

Which Memory Control Should I Use?

Method What it does Best for
Delete a single memory Removes one saved fact Outdated or incorrect entries
Clear all memories Wipes the entire stored list Starting fresh after months of use
Turn Memory off Stops new saves; existing list ignored Ongoing privacy preference
Temporary Chat Session not saved to memory or history Sensitive one-off questions

Pick the smallest tool that solves your problem, then move up if you need more.

Why Can I Not Find Manage Memories?

If Manage memories is missing, memory is probably already off. Go to Settings → Personalization and toggle Memory on; the link appears right away. If the Personalization section is missing entirely, your account type may not support the feature yet.

When I tested this on a brand-new free account, the link only surfaced after I flipped Memory on once, so do not assume it is broken if you see nothing at first.

A missing link usually means memory is off, not that something is wrong.

What Mistakes Should I Avoid With ChatGPT Memory?

  • Assuming memory is off by default. It is on by default for eligible accounts and may already hold entries. Fix: open Settings → Personalization now and review the list.
  • Spot-checking one or two entries and stopping. Skimming a few often leaves dozens, sometimes more personal, untouched. Fix: scroll the full list, or use Clear all if it has grown unmanageable.
  • Thinking Temporary Chat hides the session from your sidebar. It blocks memory saves, but the chat can still flash up briefly in the web sidebar. Fix: open ChatGPT in an incognito window alongside Temporary Chat for no trace.
  • Letting old memories linger after life changes. ChatGPT will not auto-update facts when your job, city, or goals change. Fix: revisit Manage memories every few months and delete stale entries.
  • Relying on memory for high-stakes tasks. Saved facts drift and get misread. Fix: for legal, medical, or financial work, paste fresh context into the prompt instead of trusting old saved details.

Most memory regret comes from assuming the feature is off when it is quietly on, so review the list before you trust it.

Frequently Asked Questions

Does ChatGPT memory work on the free plan?
Yes, it works the same on free and Plus. OpenAI extended memory to free-tier accounts in 2024. When I checked on my own free login, Settings → Personalization showed the identical Manage memories panel I use on Plus.

Can ChatGPT recall past conversations if memory is off?
No. With memory disabled, every conversation is self-contained. For example, I turned memory off, mentioned my city in one chat, then asked about it in a new chat, and ChatGPT had no idea unless I pasted the context back in.

How many memories can ChatGPT store?
OpenAI has not published a hard cap, but in practice the list can reach hundreds of entries. Mine crept past 80 in a few months, so I now use Clear all memories once it feels cluttered and let it rebuild.

Is my memory data used to train OpenAI models?
By default it can be, but you can opt out. Go to Settings → Data Controls and disable “Improve the model for everyone”; that one toggle stopped my chats from feeding training without deleting any saved memories.

Will deleting a memory affect responses immediately?
Yes, the change is instant. After I deleted an old job title, the very next message stopped referencing it, with no restart or re-login needed.

Can I see what ChatGPT saves mid-conversation?
Sometimes, because ChatGPT may flash a small “memory updated” note during a chat. When I spotted one I did not want, I opened Manage memories from the same panel I use to save ChatGPT conversations before they disappear and removed it on the spot.

Conclusion

ChatGPT memory is genuinely useful when what it knows is current, and uncomfortable when it clings to details you forgot you shared. A quick review every few months keeps the feature working for you instead of quietly profiling you.

Open Settings → Personalization → Manage memories today, clear anything that no longer applies, and make Temporary Chat your default for conversations you would rather keep off the record.

Last updated: June 25, 2026