Custom instructions can reduce repeated setup, but the products do not all implement them in the same way. ChatGPT, Claude, Gemini, and Microsoft 365 Copilot now offer different combinations of account-wide instructions, memory or chat-history personalization, project context, and temporary chats. Menu names, plan access, regional eligibility, managed-workspace controls, and data handling change, so use each vendor’s current settings page rather than assuming exact parity.
A short setup can save repeated typing on routine tasks. It will not make every response better: stale, contradictory, overbroad, or sensitive instructions can make responses worse and expose information more widely than intended.
This article walks through how to do it, what to put in, and the most common mistakes.
Custom instructions are for stable, low-risk preferences. Memory is for facts you are comfortable storing. Do not use either one as a place to keep customer data, credentials, confidential project names, or sensitive personal details.
Where to find the settings
In ChatGPT on web or desktop: open Settings → Personalization, enable customization, and enter instructions in the Custom Instructions field. OpenAI documents one field in the current interface, not the historical two-box layout. Current Memory controls are also under Settings → Personalization. OpenAI’s updated Memory experience can build an automatically updated memory summary from chats, files, and connected apps; some accounts may still expose the legacy saved-memory controls. See OpenAI’s current custom-instructions guide and Memory FAQ.
In Claude: inspect Instructions for Claude in Settings for account-wide guidance, and use Project instructions and knowledge for scoped work. Anthropic’s current documentation says Projects are available to all users, while enhanced knowledge, collaboration, sharing, and managed-workspace behavior depend on the plan and organization. Confirm current behavior in Anthropic’s personalization guide and Projects documentation.
In Gemini Apps: check which instructions, past-chat memory, and Connected Apps personalization your account offers. Google’s personalization guide requires a personal Google Account for the features it describes and excludes work, school, and supervised accounts; availability also differs by feature. Do not generalize this to every Gemini feature or to Gemini inside Workspace. Review the Gemini personalization guide and Gemini Apps Privacy Hub before connecting private data.
In Microsoft 365 Copilot Chat: Microsoft’s guide describes Copilot Memory, chat-history personalization, custom instructions, and temporary chat. Confirm the controls available in your work account with your administrator. A temporary chat can still be retained under the organization’s retention policy and accessible to its IT administrator during that period. This is work-account guidance, not a promise about consumer Copilot. See Microsoft’s Copilot personalization guide.
For the rest of this article we will use ChatGPT’s structure as a default, but the same principles apply across all tools.
A two-part writing structure
The current ChatGPT interface uses one Custom Instructions field. You can still organise the text into two labelled parts:
- Context: low-risk facts that materially affect the work.
- Response defaults: style, structure, uncertainty, and review preferences.
This split is useful because the two answer different questions. The first is facts. The second is style.
Part 1: About me
The goal here is to give the model the context it would otherwise have to keep asking you for. Some things worth including:
- Your role and field. “I am a senior product manager at a B2B SaaS company in Tallinn, Estonia.”
- Your professional context. “Our customers are mid-market e-commerce companies in northern Europe.”
- What you spend time doing. “Most of my AI use is around writing specs, analysing customer interviews, planning roadmaps, and drafting team comms.”
- Your background, if relevant. “I studied design and was a designer for five years before switching to product.”
- What languages you work in. “I write mostly in English; occasionally in Estonian; my Estonian is good but not native.”
- Constraints. “Use fictional examples without real customer data. A reminder from the assistant does not replace my own data check.”
Start with a few focused sentences and test them. There is no universal length that guarantees compliance; remove anything irrelevant, sensitive, or contradictory.
A useful template you can adapt:
I am [your role] at [your company / industry] in [location]. My work involves [main activities]. I have a background in [relevant prior experience]. I write in [languages]. I would describe my style as [adjective, adjective]. I care about [a thing or two you care about].
Part 2: How to respond
The response-defaults section is where you state how you normally want the model to respond. Some useful patterns:
- Length and structure defaults. “Default to short responses. Use bullet points and short paragraphs. Skip preambles.”
- Tone. “Be direct and skip filler. Preserve necessary uncertainty, correct mistakes plainly, and match my conversational tone.”
- Things to avoid. “Do not use the words ‘leverage,’ ‘utilize,’ ‘going forward,’ or ‘in today’s fast-paced world.’ Do not begin responses with ‘Great question.’”
- Calibration. “When you’re not sure, say so explicitly. Tag uncertain facts with [unclear]. Don’t fill in gaps with confident guesses.”
- Specific habits you want. “If I ask about a tradeoff, present both sides before recommending one. If I ask about a decision, ask me clarifying questions first if anything important is missing.”
- What to do with code or technical content. “When I share code, point out bugs and security issues directly. Don’t add comments explaining what the code does; assume I can read it. Suggest improvements but don’t rewrite unless I ask.”
A starting template to test:
Respond in clear, plain English. Skip unnecessary preambles (“Great question,” “I hope this helps”). Default to short replies unless I explicitly ask for depth.
Tone: direct and slightly informal. Use short paragraphs; add lists when they help. Correct mistakes plainly and preserve necessary uncertainty. Match my register if I write casually.
When you are uncertain about a fact, say so explicitly. Mark uncertain claims with [unclear]. Do not fabricate sources.
When I ask about a decision or tradeoff, ask a clarifying question if a missing detail would materially change the recommendation. Otherwise compare the relevant options and explain your recommendation and its limitations. Do not invent a numerical confidence score.
When you draft writing for me, preserve my facts and intent. Offer variants only when useful. Add a next step only if the task calls for one; do not invent commitments or requests.
Save it and test it on a low-risk task. If a default harms the answer, revise or remove it rather than assuming that a longer instruction will fix it.
ChatGPT Memory
Memory is separate from custom instructions. OpenAI’s current default experience, when enabled and available, can automatically synthesize useful context from chats, files, and connected apps into a Memory Summary. The summary is a high-level view, not necessarily a complete inventory of everything that can shape personalization. The legacy saved-memories experience remains available on some accounts and stores its memory notebook separately from chat history.
The pros: less re-explaining, more personalised responses.
The cons: it can retain or reconstruct context you did not intend to reuse, and a summary may not expose every relevant source. Recheck the settings when changing accounts instead of assuming your setup follows you.
A useful approach:
- Enable memory only if its benefit is worth retaining that context under your account’s rules.
- Review Settings → Personalization → Memory and the source indicators on personalized responses. Correct or remove sensitive, outdated, or accidental context, but do not assume the summary lists every source.
- Put explicit, stable guidance in Custom Instructions. If you ask ChatGPT to remember a fact, confirm that it appears as intended instead of treating the request as an immutable database write.
- Choose a non-personalized Temporary Chat to exclude memory, custom instructions, and plugins. Personalized temporary chats can use those sources but create no memories. Saving converts either mode into a regular chat with your account’s settings. See the Temporary Chat FAQ and privacy guide for retention and safety limits.
- To remove something fully, follow the current deletion guidance. That can mean removing it from the summary, chats and files, and disconnecting apps that can supply it. Turning memory off does not delete old chats; turning it back on can rebuild context from them.
If you are highly privacy-conscious, consider leaving memory off. Custom instructions remain stored account data, so they are not a place for secrets. Memory controls, training preferences, deletion, and authorization to use work data are separate decisions. Temporary Chat is not a guarantee of zero retention or permission to share confidential information.
What belongs where
Use this split:
| Context type | Where it belongs | Example |
|---|---|---|
| Stable response preferences | Custom instructions | ”Use short paragraphs and flag uncertainty.” |
| Stable non-sensitive personal context | Optional memory | ”I prefer metric units.” |
| Project-specific context | Project/workspace instructions | ”This project is for our public AI training site.” |
| Sensitive work details | Only a tool and workflow approved for those specific data; otherwise keep them out | Customer issue, contract clause, internal draft |
| One-off private details | Keep them out unless sharing is necessary and the tool’s terms and controls are appropriate | Personal medical, legal, family, or financial detail |
The table is a placement guide, not permission to upload. A project name, work login, or temporary-chat label alone does not establish approval for the data. Prefer fictional examples when practising.
Claude instructions and Projects
Claude provides account-wide Instructions for Claude as well as project-scoped instructions. A Project is a self-contained workspace with its own chat history and knowledge base; project instructions apply only to chats in that project.
The pattern: use account-wide instructions only for safe defaults, create a project for an approved and coherent area of work, give it the minimum necessary context, and verify current memory and knowledge behavior. Similar “Project” labels across products do not imply identical context sharing or retention.
This is useful when you switch between unrelated domains. Keep personal and client material separate, and check project visibility, member permissions, and shared knowledge before adding files. A project is an organizational tool, not proof of a security boundary. Maintaining the intended separation remains your responsibility.
Common mistakes
A few patterns to avoid.
Instructions that are too long. Long instructions become hard to maintain. Start with the few defaults that repeatedly matter and test them. Check the field limit in your current account; capacity is not a recommendation to fill it. Never remove a necessary safety rule just because the assistant fails to follow it: stop that use or change the workflow.
Stacking contradictory instructions. “Never express uncertainty” conflicts with “flag uncertain claims.” State which requirement takes priority: factual accuracy and necessary caveats come before brevity or tone. Being warm and direct need not conflict.
Instructions that lock you in. “Always respond in Estonian” or “Always use bullet points” can be annoying when you actually want English prose. Make defaults rather than absolutes: “default to bullet points unless the task is conversational.”
Forgetting where they apply. Account-wide defaults can affect unrelated conversations, while project and temporary-chat modes can change which instructions apply. For a birthday card, say: “For this draft, use a warm, conversational style.” Check the actual result rather than assuming the override worked.
A useful experiment
Use fictional source notes: “The workshop is on Friday. The venue is undecided.” Apply the default “Do not invent missing facts” and ask: “Write a short invitation using only these notes.”
Example response generated by Codex for this article on 6 September 2026:
Join us for the workshop on Friday. The venue has not been decided yet.
The response preserves the missing venue rather than inventing an address or a promise to send it later. This is an AI-generated worked example, not evidence that saved ChatGPT settings were executed or will be followed consistently. In your own test, reject an invented venue, time, or commitment.
Pilot custom instructions with low-risk preferences. When context repeats, decide whether it is stable, safe to retain, and actually useful before moving it into global instructions.
Review the resulting behavior and what the account stores. Keep a calendar reminder appropriate to how often your role, projects, and product settings change; no fixed review interval or durable quality gain is guaranteed.
Start small and review it
Add two or three low-risk defaults, open a new chat, and test them on a real task. Then ask: did the instruction help, did it leak irrelevant context into the answer, and would you be comfortable with the provider retaining it under your plan’s terms? Review memory and instructions when your role, project, or privacy requirements change.
Product guidance as of 6 September 2026
OpenAI’s Custom Instructions documentation, Memory FAQ, Temporary Chat FAQ, and Data Controls FAQ; Anthropic’s personalization and Projects documentation; Google’s Gemini personalization guide and Gemini Apps Privacy Hub; and Microsoft’s Copilot personalization guide.
