Skip to content
explore
ChatGPT logo

ChatGPT

Visit ChatGPT

Conversational AI for writing, coding, research, files, images, and automation. Capabilities and limits vary by plan.

WritingCodingResearchAgents

How it works / How to use

Conversational AI for writing, coding, research, files, images, and automation. Capabilities and limits vary by plan. State the task clearly, provide only the context that matters, and name the outcome and format you want. Pick the right surface — chat, Search, Deep research, Images, uploads, or agent tools — then refine in the same thread when the first answer is close.

  1. Open ChatGPT and start a new chat (or open a Project if the work spans multiple sessions).
  2. Choose the surface that fits: chat for drafts and explanation; Search for quick, current facts; Deep research for complex or multi-step research; Images for create/edit; uploads for your files; Think or higher reasoning on plans where it is available; Agent, Codex, or scheduled tasks when your plan includes them.
  3. Give the necessary context up front: goal, audience, constraints, source boundaries, and anything ChatGPT must not invent.
  4. Review the first reply, verify important facts and citations when accuracy matters, then refine with a targeted follow-up instead of restarting from scratch.

General chat and question answering

Lead with the question, then add context and the output shape you want (short answer, steps, pros/cons).

What to provide

  • The question or decision you need help with
  • Any background facts ChatGPT should treat as true
  • How certain the answer needs to be (quick take vs careful explanation)

Details that improve the result

  • Name your level of expertise so the explanation matches it
  • Say what to optimize for (speed, accuracy, simplicity)
  • Ask ChatGPT to flag uncertainty instead of guessing

Example prompt

Explain how DMARC, SPF, and DKIM work together for email deliverability. Audience: a developer who has never set up DNS records. Output: 5 bullets plus one common misconfiguration to avoid. If something is plan-specific or uncertain, say so.

If the first output is not good

  • Ask for one missing piece only, e.g. “Make the DNS example concrete for one provider.”
  • Request a shorter or longer version without repeating the whole prompt
  • Ask “What would you need from me to answer this more precisely?”

Common mistakes

  • Asking several unrelated questions in one message
  • Treating the first answer as final without checking obvious facts
  • Leaving out constraints that matter (deadline, region, budget, tone)
More details

Writing and rewriting

State the writing job first, then paste material under a clear label such as “Draft:” or “Source notes:”.

What to provide

  • The draft, outline, or source notes (paste in chat or upload a file)
  • Audience, purpose, and channel (email, blog, product copy, message)
  • Length, tone, and words or claims to avoid

Details that improve the result

  • Include one example of voice you like
  • List facts that must appear verbatim
  • Say whether ChatGPT should rewrite, shorten, expand, or fix only grammar

Example prompt

Rewrite this product-update email for existing customers. Tone: clear and direct, not salesy. Length: under 180 words. Keep the release date and pricing exactly as written. Do not add metrics.

Draft:
[paste draft]

If the first output is not good

  • “Keep the structure, make the opening more concrete.”
  • “Shorten by 30% and remove adjectives.”
  • “Give me two alternate subject lines under 45 characters.”

Common mistakes

  • Asking for polish without saying what “better” means
  • Forgetting to say what must not be changed
  • Mixing multiple formats (tweet + blog + FAQ) in one request
More details

Research and information gathering

Use Search for quick, current facts with citations; Deep research for complex or multi-step research and synthesis across sources; file-bound prompts when the answer must stay inside uploaded material.

What to provide

  • The topic, scope, and what you will use the answer for
  • Uploaded documents or notes when the research should stay inside your files
  • The output format (summary, comparison table, open questions list)

Details that improve the result

  • Bound the scope (“last 12 months”, “consumer plans only”, “this PDF only”)
  • For Deep research, describe the question, desired outcome, and constraints; ChatGPT proposes a research plan you can review and modify before it runs
  • Deep research results include citations or source links — visit them when accuracy matters
  • When working from files, say which document is authoritative

Example prompt

Use Deep research: summarize the main arguments for and against remote-work productivity policies published in the last two years. Return a comparison table and a “Sources to verify” section. Do not use outside knowledge for this task.

If the first output is not good

  • “Expand only the Open question column with what evidence would resolve each item.”
  • “Compare section 4 and section 7 side by side.”
  • “Rewrite the summary for a non-legal audience.”

Common mistakes

  • Assuming ChatGPT saw a file you did not upload
  • Asking for citations without providing source material to cite
  • Requesting live or time-sensitive facts without checking them yourself
More details

Image generation

Open Images or ask ChatGPT to create an image, then describe one scene with concrete visual details and exclusions.

What to provide

  • Subject, setting, style, and composition
  • Aspect ratio or use (icon, slide, poster, social post)
  • Text to appear in the image, if any, spelled exactly

Details that improve the result

  • Name camera angle, lighting, and color palette when they matter
  • Say “no text” if you do not want labels or watermarks
  • Images with thinking is available on Plus, Pro, and Business for more complex layouts

Example prompt

Create a flat illustration of a home office desk at night. Warm lamp light, laptop closed, notebook and mug, deep blue background, minimal style, no text, no watermark.

If the first output is not good

  • “Keep the same scene; change only the lighting to morning.”
  • “Make the background transparent.”
  • “Regenerate with simpler shapes and fewer objects.”

Common mistakes

  • Combining several unrelated scenes in one prompt
  • Vague mood words with no visual anchors
  • Expecting exact brand logos or trademarked characters without reference images
More details

Image understanding and analysis

Upload the image, state the analysis goal, and ask for a structured response (list, table, numbered findings).

What to provide

  • The image file (upload in chat or select a generated image)
  • What you want extracted (text, objects, issues, layout feedback)
  • Any domain context (UI review, receipt, diagram, screenshot)

Details that improve the result

  • Crop or mark the area of interest in your message if the whole image is noisy
  • Say whether you need literal transcription or interpretation
  • For screenshots, name the product surface you care about

Example prompt

Analyze this checkout screenshot. List usability issues in order of severity. For each issue: what you see, why it matters, and one concrete fix. Ignore marketing copy above the fold.

If the first output is not good

  • “Focus only on mobile spacing and tap targets.”
  • “Transcribe all visible text exactly.”
  • “Turn the findings into a QA checklist.”

Common mistakes

  • Uploading a blurry image and asking for fine text extraction
  • Asking for medical, legal, or safety decisions from a photo alone
  • Not saying whether you want description or critique
More details

Image editing

Select the image, describe the edit as a delta (“remove background”, “change jacket color to navy”, “add soft shadow”), not a full re-description of the whole scene.

What to provide

  • The source image (generated in ChatGPT or uploaded)
  • The single change or small set of changes you want
  • Elements that must stay the same

Details that improve the result

  • One major edit per pass works better than many at once
  • Name what must not move (face, logo placement, composition)
  • Regenerate from the closest version instead of starting over

Example prompt

Edit this image: remove the person on the left, keep the street and signage unchanged, preserve the same color grading.

If the first output is not good

  • “The background is right; fix only the hand artifact.”
  • “Make the edit more subtle.”
  • “Try a version with a transparent background.”

Common mistakes

  • Asking for a completely different scene while calling it an edit
  • Stacking unrelated edits before checking the first result
  • Forgetting to say which parts are locked
More details

File and document analysis

Upload up to the per-message file limit, name each file, and state the operation before asking follow-ups.

What to provide

  • The file (PDF, DOCX, spreadsheet, presentation, code, or text)
  • The task: summarize, compare, extract, transform, or analyze
  • Rules for quotes, numbers, and what must not be invented

Details that improve the result

  • For spreadsheets, say which columns or sheets matter
  • Ask for counts, quotes, or headings when you need extraction
  • Note plan limits on uploads and file size (see OpenAI file-upload help)

Example prompt

I uploaded sales_q1.csv. Calculate total revenue by region, show the top 3 regions, and list any rows with missing region values. Return a small table plus one sentence of interpretation. Do not chart yet.

If the first output is not good

  • “Now chart only the top 3 regions as a bar table.”
  • “Quote the exact sentence from page 2 that supports this summary.”
  • “Compare file A and file B on pricing terms only.”

Common mistakes

  • Uploading many files without saying how they relate
  • Asking ChatGPT to invent numbers missing from the file
  • Expecting perfect layout extraction from complex PDFs without verifying against the source file
More details

Coding

Describe the task, paste relevant code or error output, and ask for code plus a brief explanation or usage example.

What to provide

  • Language, framework, and runtime
  • The exact behavior or bug to address
  • Inputs, outputs, edge cases, and style constraints

Details that improve the result

  • Include the error message verbatim
  • Say whether you want a patch, a full function, or a review
  • Codex and Agent flows on supported plans can work across multi-step coding tasks

Example prompt

In TypeScript, write a function `parseIsoDate(value: string): Date` that accepts YYYY-MM-DD only and throws a typed error for invalid input. Include one passing and one failing usage example. No external libraries.

If the first output is not good

  • “Handle empty string explicitly.”
  • “Rewrite using Zod instead.”
  • “Explain the failure case in two sentences.”

Common mistakes

  • Omitting the language or framework
  • Pasting huge files without pointing to the relevant section
  • Accepting code without running or reading it
More details

Code checking, debugging, and explanation

Paste the error first, then the smallest code sample that reproduces it, then ask for diagnosis steps and a fix.

What to provide

  • The code snippet or error log
  • What you expected vs what happened
  • Environment details that matter (browser, Node version, OS)

Details that improve the result

  • Include stack traces and line numbers when you have them
  • Say whether you want a fix, a root-cause explanation, or both
  • Ask for a test that would catch the bug

Example prompt

This React component re-renders infinitely in development. Expected: fetch runs once on mount. Actual: fetch loops. Here is the component and the console error:
[paste code + error]
Explain the cause, propose a minimal fix, and note any trade-offs.

If the first output is not good

  • “Show the fix as a diff against my snippet.”
  • “Explain like I’m new to React effects.”
  • “Suggest a regression test.”

Common mistakes

  • Sharing only the error string with no code
  • Changing several things at once after the first fix
  • Assuming ChatGPT can see your private repo without pasted context
More details

Structured output and formatting

Say “Return only …” and define columns or keys explicitly. Ask for valid JSON when you will paste the result elsewhere.

What to provide

  • The raw material (notes, bullets, chat history, or file)
  • The target format (table, JSON, checklist, outline, CSV-style rows)
  • Field names, column order, and empty-value rules

Details that improve the result

  • Specify max length per field
  • Say whether nulls, em dashes, or “N/A” are allowed
  • Request markdown tables when humans will read the output

Example prompt

Turn these meeting notes into a markdown table with columns: Owner, Task, Due date, Blocker. Use “TBD” only when a date is missing. Do not add tasks that are not in the notes.

Notes:
[paste notes]

If the first output is not good

  • “Convert the same content to JSON with keys owner, task, dueDate, blocker.”
  • “Sort by due date ascending.”
  • “Remove the Blocker column and add Priority.”

Common mistakes

  • Ambiguous column names
  • Not saying whether extra fields are allowed
  • Asking for JSON without defining types or required keys
More details

Tasks and automation (where your plan supports them)

For scheduled tasks, describe the trigger, inputs, and output template. For Agent/Codex work, state the goal, allowed tools, and stop conditions.

What to provide

  • The recurring goal and schedule (scheduled task availability and active-task limits vary by plan)
  • Inputs ChatGPT should use each run
  • The deliverable format and what requires human approval

Details that improve the result

  • Keep automations focused on one outcome
  • Say what must never happen without confirmation (send email, purchase, publish)
  • Review the first run before relying on a schedule

Example prompt

Schedule a weekly task: summarize new items from my project notes file and return 1) three priorities 2) blockers 3) suggested next message to the team. Do not send anything externally. Review the first run before relying on the schedule.

If the first output is not good

  • “Shorter summary, bullets only.”
  • “Add a section for risks I missed.”
  • “Run once now as a test using yesterday’s notes.”

Common mistakes

  • Automating vague goals (“keep me updated on everything”)
  • Granting irreversible actions without review
  • Assuming a feature is on your plan without checking ChatGPT settings
More details

How to prompt

ChatGPT responds best when you treat each message as a brief: task, necessary context, desired outcome, output format, and constraints. Use the same thread to iterate — targeted follow-ups usually work better than rewriting from zero.

Goal: draft a support reply.
Context: customer cannot reset password; account email is confirmed.
Output: 3 short paragraphs + numbered steps.
Constraints: empathetic tone, no refund offer, under 120 words.

Name audience and tone

Say who the output is for and what tone to use — formal, concise, friendly, or professional. OpenAI guidance favors clear, specific prompts with preferred tone and style.

Example

Explain eigenvalues to a first-year engineering student. Tone: patient and plain. Use one everyday analogy and one tiny numeric example.

Separate facts from instructions

Put source material under a labeled block and instructions above it so ChatGPT does not rewrite your facts accidentally.

Example

Use only the facts in SOURCE. Do not add statistics.

SOURCE:
[paste facts]

Task: Write a 100-word FAQ answer.

Ask for format explicitly

Name the structure you want back: table, JSON, outline, checklist, or diff. Explicit formats reduce cleanup time.

Example

Return JSON only: { "title": string, "steps": string[], "risks": string[] }. No markdown fences.

Iterate with targeted follow-ups

Reference the previous answer and change one dimension at a time: shorter, more formal, more examples, or fix one paragraph.

Example

Keep paragraphs 1 and 3. Rewrite paragraph 2 to mention migration downtime in plain language.

Pick Search or Deep research when facts must be current

Use Search for quick, current facts with links. Use Deep research for complex or multi-step research and synthesis. Deep research proposes a research plan you can review and modify; results include citations or source links.

Example

Search: What changed in EU digital-markets enforcement in the last 90 days? Cite sources.

Deep research: Compare remote-work productivity findings from the last two years. Return a table and list open questions.

Use Think or higher reasoning for harder problems

On Free and Go, select Think for harder questions when it is available in your app. On eligible paid plans, you can choose a higher reasoning level when a task needs deeper planning or analysis.

Example

Think through this step by step: we have 12 support tickets with overlapping billing bugs. Propose a triage order and the first three questions to ask the team.

Use uploads and Images intentionally

Upload files when the task depends on your data. Switch to Images when the task is visual creation or editing—not plain chat.

Example

Upload budget.xlsx, then: “Summarize variance by department and flag any line item over 10% vs plan.”

Best output tips

Lead with context, goal, and constraints

State what you are trying to do, who it is for, and any limits (length, tone, deadline, region, budget) before the detail. ChatGPT fills gaps when these are missing.

Name the output format explicitly

Ask for bullets, a table, JSON, a checklist, steps, or a specific word count. Explicit structure reduces cleanup and keeps follow-ups focused.

Paste examples or references when tone matters

One sample paragraph, subject line, or UI label often communicates voice better than adjectives like “professional” or “friendly”.

Separate facts from instructions

Put source material under a labeled block (SOURCE, Draft, Notes) and say which facts must appear verbatim and which must not be invented.

Ask ChatGPT to flag uncertainty

Phrases like “List assumptions”, “Mark anything you are not sure about”, or “Separate verified facts from open questions” reduce confident-sounding guesses.

Iterate with targeted follow-ups

Reference the previous answer and change one thing: shorten, fix one paragraph, add an example, or adjust tone. Avoid restarting from scratch when the draft is close.

Break complex work into steps

For multi-part tasks, ask for a plan first or work one section at a time. Confirm each step before moving on—especially for code, research, and long documents.

Pick the right ChatGPT surface

Plain chat for drafts and explanation; Search for quick, current facts with citations; Deep research for complex or multi-step synthesis; Images for create/edit; uploads for your files; Think or higher reasoning on plans where available; Agent, Codex, or scheduled tasks when your plan includes them.

For files and documents

Upload first, name each file, state the task, point to the sections or columns that matter, and say the output you want back (summary, table, extraction, comparison).

For image generation

Describe subject, purpose, composition, important visual details, style when relevant, and aspect ratio or orientation. Say “no text” or spell any text exactly when labels matter. Images with thinking is available on Plus, Pro, and Business for more complex layouts.

For image understanding and editing

Upload or select the image, state whether you need transcription, critique, or an edit, and name what must stay unchanged (face, layout, color grading, background).

For coding and debugging

Include relevant code, the exact error or output, expected vs current behavior, environment (language, framework, runtime), constraints, and what should not be changed.

For research and verification

Use Search for quick, current facts with links. Use Deep research for complex or multi-step synthesis; review and modify the proposed research plan before it runs. Visit citations or source links when accuracy matters. Treat output as a first draft, not a final source.

For structured outputs

Define column names, key order, empty-value rules, and whether extra fields are allowed. Request valid JSON when you will paste the result into another tool.

For tasks, automation, and Agent/Codex

Keep each automation focused on one outcome, list inputs and deliverable format, and require human approval before irreversible actions (send, purchase, publish). Scheduled task availability and active-task limits vary by plan.

For voice and dictation

Use Voice for live back-and-forth conversation. Use Dictation when you want to record speech, review and edit the transcription, then send it as text. Include exact dates or time zones for time-sensitive questions and verify important answers afterward.

For custom GPTs

All signed-in users can use GPTs they have access to. Creating or editing GPTs requires an eligible paid subscription; in Business, Enterprise, or Edu workspaces, workspace settings and permissions also apply.

Use Projects for ongoing work

When a task spans multiple sessions, a Project keeps related chats and files together so follow-ups stay in context.

Constrain length early

Word counts, bullet limits, and “one paragraph only” prevent overlong first drafts you still have to trim.

Verify numbers and quotes

ChatGPT can summarize files you provide, but confirm figures, dates, and quotations against the source before sharing or acting on them.

  • State the task, necessary context, desired outcome, format, and constraints before the detail — right-size context instead of dumping everything.
  • Start new chats for unrelated tasks; keep one thread when you are refining the same deliverable.
  • Use Search for quick, current facts; Deep research for complex multi-source synthesis; review the Deep research plan and verify citations when accuracy matters.
  • For files and images, upload first, then refer to “the uploaded file/image” in your prompt.
  • Say what ChatGPT must not invent: dates, prices, metrics, quotes, or policy claims.
  • When an answer is close, ask for a diff-style change instead of regenerating everything.
  • Break complex work into steps, or use Deep research / Think when the task needs planning or synthesis.
  • Check plan-specific features (Images, file limits, scheduled tasks, Agent/Codex) before relying on them.
  • Treat ChatGPT output as a first draft — verify numbers, quotes, code, and policy language before using them.

Try this AI

Try ChatGPT

Product Details

Pricing, features, limits and latest updates

ChatGPT

Conversational AI for writing, coding, research, files, images, and automation. Capabilities and limits vary by plan.

Free / Paid · Free

Pricing Plans

Free

Free

Go

$8/ Monthly

Plus

$20/ Monthly

Pro $100

$100/ Monthly

Pro $200

$200/ Monthly

Business Standard

$25/ Monthly

Business Standard

$240/ Yearly

Business Premium

$125/ Monthly

Business Premium

$1,200/ Yearly

Enterprise

Custom

Key Features

Image Generation

ChatGPT Images 2.0 is available on all tiers; Images with thinking for complex layouts is on Plus, Pro, and Business. Free is limited and slower (OpenAI Images help).

Chat

Conversational assistant on web, iOS, Android, and desktop. Projects group related chats, files, and instructions; Library stores uploaded and created files on web. Limits vary by plan.

Voice

Voice for live conversation; Dictation for spoken input you review as text before sending. Verify time-sensitive answers and check transcripts afterward (OpenAI Voice help).

API

The OpenAI API is billed and managed separately from ChatGPT subscriptions (OpenAI Help Center).

Automation

Scheduled tasks on Go and higher (active-task limits vary by plan). All signed-in users can use existing GPTs; new GPT creation and publishing are not available on personal Free, Go, Plus, or Pro accounts (OpenAI Help Center).

Agents

ChatGPT Agent, Codex, and ChatGPT Work run multi-step tasks from supported surfaces; availability and monthly limits vary by plan (OpenAI Help Center).

Image Editing

Upload or select an image and describe edits as a delta; use the Select tool for region edits (OpenAI Images help).

Cited Search

Search returns quick web summaries with citations; Deep research produces documented multi-source reports. Use Search for quick facts and Deep research for depth (OpenAI Help Center).

Limits

  • Max file size: 512 MB. Hard limit per file in chats, GPTs, and Projects (OpenAI File Uploads FAQ).
  • Free file uploads: 3 per day. Free-plan daily file upload allowance (OpenAI File Uploads FAQ).
  • Rolling upload rate: 80 files per 3 hours. Rolling upload rate cap; may be lowered during peak hours (OpenAI File Uploads FAQ).
  • Image upload size: 20 MB. Maximum image upload size (OpenAI Library help).
  • Spreadsheet size: 50 MB. Approximate maximum CSV/spreadsheet upload size (OpenAI Library help).
  • Library storage: Free 500 MB; Go 4 GB; Plus and Business 20 GB; Pro 100 GB (OpenAI Library help).
  • Project file storage: Per-project file caps vary by plan; OpenAI sources disagree on Plus limits—check in-product limits (Projects and File Uploads help).
  • Simultaneous uploads: 10 files. Maximum files uploaded at the same time in Projects (OpenAI Projects help).
  • Scheduled task frequency: 1 per hour. Tasks cannot run more than once per hour (OpenAI Scheduled Tasks help).
  • Active scheduled tasks: Go up to 3; Plus up to 5; Business and Edu up to 10; Pro and Enterprise up to 15 (OpenAI Scheduled Tasks help).
  • Agent mode messages: Plus 40/month; Pro 400/month; Business 40/month per initial user-initiated request (OpenAI ChatGPT agent help).
  • Deep research usage: Allowance varies by plan; check the in-product usage counter (OpenAI Deep research help).
  • OpenAI API billing: OpenAI API usage is billed separately from ChatGPT subscriptions (OpenAI Help Center).

Ideas / Prompt experiences

Share a prompt that worked for you. Username and email are shown with your submission. External links are not allowed.

Example prompt

Support reply with constraints

Prompt

Draft a support reply for a customer who cannot reset their password. Tone: empathetic, direct. Under 120 words. Keep the steps numbered. Do not offer a refund. Context: account email is confirmed; reset emails are not arriving.

Variation

Keep paragraphs 1 and 3. Rewrite paragraph 2 to mention checking spam and whitelisting the sender domain.

Short explanation

OpenAI guidance: state the task, necessary context, tone, format, and constraints—then iterate with targeted follow-ups.

Example prompt

Deep research with source verification

Prompt

Use Deep research: summarize the main arguments for and against remote-work productivity policies published in the last two years. Return a comparison table and a separate “Sources to verify” section.

Short explanation

Use Deep research for complex multi-source synthesis; Search is better for quick facts. Review and modify the proposed research plan before it runs.

Share your experience

Required fields are marked. Variation, result, and explanation are optional.

You may also like

Claude logo

Claude

Explore

AI assistant for writing, analysis, coding, and collaborative problem-solving.

WritingFree / Paid

Perplexity logo

Perplexity

Explore

An answer engine that searches the web and cites its sources.

WritingFree / Paid

Notion AI logo

Notion AI

Explore

AI assistant in Notion for agents, enterprise search, meeting notes, and workspace writing.

WritingFree / Paid

Back to Explore