Company context for an AI assistant: what to write and where to put it

What goes on the page

1. Role and goal, in two sentences.
2. Audience and tone: who reads the output, and two or three tone rules.
3. Can and can't: each "can't" says what to do instead.
4. Sources of truth: which document answers which topic, and what to do when nothing does.
5. Format: the shape of a good answer, with one or two examples.
6. Edge cases and hand-off: the three to five situations most likely to go wrong, and when to pass to a person.

Where to put it

A worked example

Illustration: a made-up business. A 20-person company sells booking software to physiotherapy clinics. Its support lead, Tomás, writes the context for the team's support project:

Role and goal: You help our support team answer clinic owners and
front-desk staff. A good answer is correct, cites the help-center
article, and is ready to paste.

Audience and tone: Busy clinic staff, not technical. Plain words,
short steps, no exclamation marks.

Can: explain features, settings and billing from the sources below.
Can't: promise refunds or release dates; say "I'll check with the
team" and tag the support lead. Don't give clinical advice; say we
can't help with that.

Sources of truth: help center for how-to; the pricing page for plans;
the status page for outages. If they disagree, the newest wins and you
flag it. If nothing answers, say so and suggest a ticket.

Format: the answer in one line, then at most five numbered steps,
then the article link. Examples: three past tickets and the answers
we sent, in the project file examples.md.

Hand off: angry or legal messages, data deletion requests, anything
about a patient's records.

Tomás tested it on ten real tickets. Two answers promised a fix date, so he added the release-date line; everything else came from the first draft. The page stays under 25 lines because each line traces to a real mistake or a likely one.

Write yours

1. Collect ten real requests the assistant will get.
2. Write the six parts above, shortest first draft you can.
3. Point to documents; don't paste them in.
4. Test on the ten requests and add a rule only where an answer went wrong.
5. Put it where it's read on its own: a project, a skill, or the repository file.

Next step: the plan your team shares it on

Context that one person keeps in a personal account helps one person. To share it with the team, you need a team plan: ChatGPT plans for a small business covers which ChatGPT plan fits, and ChatGPT Business vs Claude Team compares the two team plans. For the idea behind all this, see context engineering.

Questions

What company context should I give an AI assistant?

One short page: who you serve and what a good answer does for them, your products and your own words for things, what the assistant must never promise and what to say instead, which document is the source of truth for which topic, the format you want with an example, and when to hand off to a person.

Where do I put company context for ChatGPT or Claude?

For one team or account, in a project's instructions, with long documents as the project's files. For a task the whole company repeats, in a skill. For coding agents, in the repository's CLAUDE.md or AGENTS.md.

How long should company context be?

As short as it can be while still changing the answers. Anthropic's engineering team describes good context as the smallest set of high-signal information; point to long documents instead of pasting them in, and add a rule only when a test shows the assistant needs it.