ChatGPT now covers several kinds of work, but the best starting point is still the task in front of you. A short question belongs in Chat. A longer assignment with a defined deliverable may fit Work. A software change belongs in Codex. Files, images, voice, web search, Projects, and memory add context, but they do not remove the need to check the answer.
This cheat sheet is a quick reference, not a catalog of every button. OpenAI changes features through staged rollouts, and availability can differ by plan, workspace, region, device, app version, and administrator settings. If a control described here is missing, check the linked OpenAI documentation and your account settings rather than assuming every interface is identical.
Choose Chat, Work, or Codex by task
OpenAI’s current ChatGPT Work and Codex guide separates the three experiences by purpose. This distinction is useful only where Work or Codex is available for your plan and workspace.
| Experience | Use it for | Give it | Check before accepting |
|---|---|---|---|
| Chat | Questions, explanations, brainstorming, rewriting, and quick web searches | A clear request, relevant context, and the output format | Facts, calculations, citations, and whether it followed the brief |
| Work | Longer research, analysis, and finished documents, spreadsheets, presentations, reports, or Sites | The desired deliverable, source material, constraints, and review criteria | Progress, permissions, source quality, and the final artifact |
| Codex | Writing or debugging code, running commands and tests, reviewing changes, and working with repositories | The repository or folder, acceptance criteria, commands to run, and files that must not change | The diff, test output, security implications, and unintended file changes |
Use Chat when conversation is the work. Choose Work when you can describe a finished artifact and want an agent to carry out several steps. Choose Codex when the assignment depends on source code, a terminal, tests, or repository history. On desktop, local access requires permission; cloud and local conversations can also follow different storage and syncing behavior. Read the official guide before granting access to folders or apps.
Text prompts: specify the job, evidence, and finish line
A useful prompt usually answers four questions: What should ChatGPT do? What context may it use? What constraints matter? What would a finished answer look like? You do not need a long template for a simple request.
Example prompt: “Rewrite the note below for a customer. Keep every date and price unchanged, use plain English, and return a subject line plus a message under 180 words. If a fact is missing, mark it as a question instead of inventing it.”
For a complicated task, separate production from review. Ask for a draft, then ask ChatGPT to compare that draft with your stated criteria. A final human check still matters, especially for material that will be published, sent to a customer, submitted for study, or used for a consequential decision.
A few everyday patterns are easy to reuse. For a meeting, provide the notes and ask for decisions, unresolved questions, owners, and deadlines without adding details. For learning, ask for an explanation followed by questions that test understanding. For planning, give the goal, deadline, available resources, and constraints, then request an ordered checklist. For editing, identify the audience and facts that cannot change. In each case, save the original beside the output so you can compare them.
Files and spreadsheets: ask for traceable work
According to OpenAI’s capabilities overview, ChatGPT can work with uploaded documents and can analyze structured data by running code in a secure environment. Availability and limits depend on the account and feature. Upload only material you are permitted to share.
- For a document, name the section, page, clause, or question you care about. Ask the answer to quote or point to the supporting passage.
- For several documents, define the comparison fields before requesting a summary.
- For a spreadsheet, identify the columns, date range, missing-value rule, and expected chart or table.
- Ask for the method, assumptions, and any rows that were excluded. Download and inspect important outputs.
Example prompt: “Using the attached CSV, group revenue by month and product. List the cleaning steps, flag missing dates, create a table, and explain any assumption before applying it.”
Do not treat a polished chart as proof that the source data was read correctly. Compare totals with the original file and spot-check several rows.
Images: describe the visual question
ChatGPT can analyze uploaded screenshots, charts, diagrams, and photos, and image generation is available in supported experiences. Tell it what to inspect rather than asking only “What is this?” For a chart, name the axis or comparison. For a screenshot, identify the error or interface region. For generated images, describe the subject, intended use, composition, text requirements, and aspect ratio.
Image reading can miss small labels, obscured details, and ambiguous objects. Crop to the relevant area, upload a clearer version, or paste the text when accuracy matters. Check generated images for spelling, brand use, misleading details, and whether they meet the publishing platform’s requirements.
Search: use it for current facts, then open the sources
ChatGPT Search can retrieve timely information and show citations. It may search automatically when a question would benefit from the web, or you can choose Search in supported interfaces. Search availability and usage remain plan dependent.
Include an exact date, location, or version when the answer is time sensitive. Ask for primary sources when possible. Then open the cited links and check that they support the sentence beside them. A citation can be real but irrelevant, stale, or too weak for the claim.
Example prompt: “Find the current official policy for this feature. Use the vendor’s documentation first, state the page’s date if shown, and separate documented facts from your interpretation.”
Voice: talk through ideas, but review the transcript
OpenAI’s current ChatGPT Voice guide describes spoken conversations on supported web, mobile, and desktop experiences. Voice options vary by plan, workspace, region, and app version. Some voice experiences can use search, memory, text, or images, while other capabilities differ.
Voice works well for brainstorming, rehearsing questions, or discussing a document while your hands are occupied. State an exact date and time zone for schedule questions. After the conversation, review the text because a transcript may not match every spoken word. Use dictation instead when you want to edit a recorded prompt before sending it.
For device-specific setup and permissions, use the site’s ChatGPT mobile guide or ChatGPT desktop app guide for Mac.
Projects: keep an ongoing job together
Projects in ChatGPT group chats, files, and project instructions around one objective. They are useful when the same assignment continues across several sessions. Start a project for a client, course, research question, or recurring report, then add only the sources needed for that work.
Project instructions apply inside that project and can override global custom instructions. Memory behavior also depends on how the project was created, the plan, sharing state, and account or workspace settings. Shared projects expose their chats and files to project members, so review the audience before adding confidential material. The dedicated ChatGPT Projects guide covers setup in more depth.
Memory and Custom Instructions are different controls
Memory can use context from chats, files, and connected apps to personalize later responses when enabled. OpenAI’s Memory FAQ explains the current memory summary and the legacy saved-memory option. The summary is a high-level view and may not display every factor used for personalization.
Custom Instructions are explicit standing directions about what ChatGPT should know and how it should respond. Put stable preferences there, such as your default language or response style. Do not use memory as the only place to store a requirement that must be followed in a specific task. Repeat that requirement in the prompt or project instructions.
Deleting a chat does not necessarily delete a saved memory created from it, and deleting a saved memory does not remove old mentions from past chats. For a careful explanation of review and deletion, read the site’s ChatGPT Memory and Controls guide. For explicit standing preferences, use the Custom Instructions guide.
Data Controls and Temporary Chat
Memory settings and model-training settings solve different problems. OpenAI’s Data Controls FAQ says that turning off “Improve the model for everyone” prevents new conversations from being used to improve models while allowing them to remain in chat history. The setting applies across the signed-in account.
Temporary Chat does not appear in history, does not use or create personalization memories, and is not used to improve models. OpenAI may retain a copy for up to 30 days for safety. Enabled Custom Instructions still apply. If a GPT action sends information to a third party, that recipient’s privacy policy governs the data it receives.
Temporary Chat is useful for an isolated conversation, but it is not permission to upload secrets, customer records, health information, or workplace material. Apply the same data-minimization rule everywhere: provide only what the task needs and what you are authorized to share.
A verification habit that works across tasks
- Restate the deliverable and constraints before accepting a long result.
- Check names, dates, quotations, calculations, and links against the original source.
- For files, compare totals and sample rows. For code, inspect the diff and run the required tests.
- Ask ChatGPT to list assumptions and uncertainty, but do not rely on its self-critique alone.
- Confirm that no private data, hidden instruction, or unintended file was included.
- Have the responsible person approve anything consequential before it is published or acted on.
This article summarizes documented product behavior and example workflows. It does not claim personal testing of every plan, device, or rollout.
Frequently asked questions
Should I use Chat or Work for a long report?
Use Chat when you want an interactive discussion or draft. Use Work, where available, when you can define a finished report, provide sources and constraints, and review progress as it completes several steps.
Is Codex the best choice for every technical question?
No. Chat can explain a concept or discuss a short code sample. Codex is better suited to repository work that involves files, commands, tests, and review of actual changes.
Does Temporary Chat turn off Custom Instructions?
No. OpenAI says Temporary Chat still follows enabled Custom Instructions, although it does not use or create memories for personalization.
Can I trust an answer because it includes citations?
No. Open each important source and confirm that it supports the specific claim. Prefer current primary documentation for product behavior, policies, prices, and availability.