How to Build Personal Context for AI: 6 Steps
What to include, what to cut, a template you can copy, and how to keep it current across every AI tool.
To build personal context for AI, write down only what changes a good answer, pull it from sources that already describe your work, and keep it as short structured fields you load into each tool. Most people write a bio instead, a paragraph pasted into a system prompt, which makes the model guess what matters and goes stale the moment your work moves on.
Good personal context is structured, current, and short enough that every line changes the answer. Here's how to build it in six steps.
What should personal context include?
Include what changes what a good answer looks like, and nothing else. That rule decides every line, so it's worth starting there.
Four kinds of information cover most of the value. Who you are (role, field, level of expertise) calibrates depth and framing; a senior engineer and a beginner should get different answers to the same question. What you're working on turns generic advice into advice for your situation. How you want answers ("direct, no preamble, bullets over prose") saves you re-requesting it every session. And the specifics a task depends on, like a client's brand rules or a codebase's conventions, belong in context only when that task is in play.
How do you build personal context, step by step?
Six steps. The first four produce the file; the last two decide whether it stays useful.
1. Sort what you know into four layers
Split your context by how fast it changes, so the fast-moving parts don't get buried next to permanent ones:
- Identity: name, role, company, location. Changes about once a year.
- Profile: background, expertise, communication style, tools you use. Changes when your role shifts.
- Mind: active projects, current priorities, decisions in progress. Changes weekly.
- Environment: the people, companies, tools and recurring meetings around you.
You don't need all four on day one. Identity and Profile are enough to make a visible difference, and they're the easiest to collect. Add Mind before you start a project, and Environment when you want help with people and scheduling.
2. Collect it from sources that already describe your work
Pull from what exists instead of writing from memory. A self-written bio drifts toward how you'd like to be seen; your actual work is more accurate.
- Identity and Profile: your LinkedIn profile, personal website, old CVs, and your GitHub activity, which shows your real stack better than a self-reported list.
- Mind: Notion or your project tool, the documents you've edited recently, your own notes on what matters this month.
- Environment: Google Calendar (who you meet and how often), Gmail or Slack (who you work with), your CRM if you have one.
Cross-check as you go. If LinkedIn says one thing about your role and Notion says another, keep the version that matches what you actually do this month.
3. Write it as short structured fields
Write a form, not a letter. Short labeled fields let a tool find the fact it needs instead of interpreting a paragraph. Here's a template, filled in for a fictional marketer:
## Identity
Name: Ana Kowalska
Role: Head of Growth
Company: Ledgerly (B2B invoicing SaaS, 40 people)
Location: Warsaw, Poland (CET)
## Profile
Background: 8 years in B2B marketing, previously demand gen at two seed-stage startups
Expertise: paid acquisition, lifecycle email, positioning
Tools: HubSpot, Notion, Linear, Claude, ChatGPT
Style: direct, short paragraphs, British English, no exclamation marks
## Mind
Active: Q4 pricing page relaunch, LinkedIn ads test for mid-market ops buyers
Priorities: cut CAC below the Q3 level, ship the pricing page before 15 November
Open decision: whether to gate the ROI calculator
## Environment
People: Tomek (CEO, approves budget), Ola (designer), Jan (sales lead)
Customers: finance teams at EU companies with 50 to 500 people
Recurring: Monday growth standup, monthly board update
Three rules keep it usable. Short fields beat sentences. Specific beats general: "paid acquisition, lifecycle email" says more than "marketing". And Mind is a snapshot, not a diary, so anything from last quarter goes.
4. Cut what the model can infer, what's stale and what doesn't matter
Cut hard, because noise isn't harmless: every irrelevant line dilutes the lines that matter. Three kinds of detail go:
- Inferable detail. A capable model knows what a Head of Growth does. State the role and spend the space on what's specific to you.
- Stale detail. The finished project, the old priority, the preference you've dropped. Outdated context is worse than none, because the model applies it with full confidence.
- True but irrelevant detail. Your hobbies rarely change how an AI should answer a work question.
The test for every line is the same: does this change the answer? Research on agent context files backs the instinct. ETH Zurich researchers found that context files an LLM generated slightly lowered coding-agent success and raised inference cost by about 20%, while files written by developers did better but still improved success only marginally (Gloaguen et al., 2026). The generated files mostly repeated what the agent could already work out. Lean and specific wins.
5. Load it into the tools you use
Put the file where each tool reads it at the start of a session:
- ChatGPT: custom instructions, which hold up to 5,000 characters on Plus, Pro, Business, Enterprise and Education plans since July 2026 (ChatGPT release notes), or a project's instructions. More in How to Save Context in ChatGPT.
- Claude: the Instructions for Claude field in Settings, or a Claude Project. See How to Set Up Preferences for Working With Claude.
- Claude Code, Codex and Cursor: a
CLAUDE.mdorAGENTS.mdin the repo for project rules. - As a skill: ChatGPT and Codex can load a skill, a
SKILL.mdfile with instructions, when a task matches its description; in ChatGPT you can also pick one with@, in the Codex CLI with$(ChatGPT Learn). - Over MCP: tools that support MCP can pull your context instead of reading a pasted copy. Setup per tool: How to Load Personal Context via MCP.
Pasting works, but every pasted copy is a separate snapshot you now have to update.
6. Keep it current
Set a schedule per layer, because each one ages at a different speed:
| Layer | When to update |
|---|---|
| Identity | When your role or company changes |
| Profile | Every few months, or when your focus shifts |
| Mind | Weekly: add new projects, drop finished ones |
| Environment | When a client, project or key colleague changes |
The failure everyone hits: Identity and Profile get set up carefully, then Mind is never touched again. A month later the AI knows who you are but not what you're doing. A five-minute weekly review fixes it, as long as you actually do it, and in every tool where you pasted a copy.
Is there a faster way to build personal context?
Yes: connect the apps where your context already lives and let them keep it current. The manual method has two costs, the first setup and the weekly discipline, and the second one multiplies with every tool you paste into.
Unabyss builds your personal context from Gmail, Google Calendar, Notion, Slack, GitHub, Linear and other apps, keeps it current as they change, and serves it to Claude, ChatGPT, Cursor and other MCP clients. Setup takes about 90 seconds: connect at least two apps and let it build. In an August 2026 Unabyss benchmark, that cut cost by 36% and time by 38% against an assistant with connected tools alone, while a hand-written CLAUDE.md barely moved either number (What Structured AI Context Is Worth).
The Mind layer still benefits from a line or two of your own, since some priorities don't live in any app. To see how the tools compare, read Best AI Memory Tools, or connect your apps to Unabyss and skip the weekly update.
