Stop repeating yourself to AI: split company info into 3 parts
Split company info for AI into fixed terms, task guidance and case records, and you stop pasting it all every time.
Key points
- Send AI only fixed terms, like pricing
- File guidance by task: quotes, replies
- Log one-off requests, not as fixed terms
- Test both reply types: price, no price
Contents
Send AI only the company terms that never change
A&A perspective
The company information you give AI is the set of your own terms and decision material that AI reads when it drafts a reply or a quote. Of that, what you send every time should be only the terms that hold for every job, such as how you set prices and what work you cover.
A&A perspective
A common habit is to paste your company profile, price list, past sales call notes and service ideas still under discussion into every request. It does save explaining. But ideas you have not settled, or a promise made to just one client, can end up in the draft as if they were your current terms.
From the sources
In a technical article on choosing what information to give AI, Anthropic writes that this information should be treated as a finite resource. It also writes that as the amount grows, the model's precision in picking out the right piece can drop. It describes this as a gradual decline rather than a sudden failure.
From the sources
On that basis, the article recommends choosing the smallest set of information that most helps the outcome you want. It adds that smallest here does not mean the same thing as short.
From the sources
Claude Code has an instruction file, CLAUDE.md, that it reads at the start of every session. Anthropic's official documentation describes this file as the place to write down what you would otherwise explain again.
From the sources
The documentation recommends keeping in this file only what every session needs, and moving procedures that matter for only part of the work into a mechanism that is read only when needed. It also gives a target length of under 200 lines.
A&A perspective
Automate & Augment (A&A), which writes this article, applies the same idea to company information. Put the terms you send every time into a single document: what the company does, how it sets prices, what work it covers, and what work it does not take on. Write them one line each, for example "We quote by working hours" or "We do not run ad campaigns."
A&A perspective
This one document also holds a table of contents that shows where the rest of your material lives.
A&A perspective
These sources are explanations written for developers who build AI. They did not measure the effect on a company's business documents.
Split guidance by task and let AI read it only when needed
A&A perspective
AI reads your quote guidance only when it drafts a quote, and the points to ask back in a reply only when it writes a reply. Splitting guidance by task lets you hand over just what this job needs.
From the sources
Anthropic describes Agent Skills, a way to give AI folders of material on how to do particular kinds of work. At first, AI reads only the name and a short description of each skill. It reads the full content only when it judges the skill relevant to the task at hand.
From the sources
The content can also point to further files that are read only in the situations that need them. Because AI uses the name and description to decide whether to read the content, Anthropic advises paying special attention to them.
From the sources
Claude Code supports these as skills. Its official documentation recommends putting procedures that matter for only part of the work into a skill. Skills load only when you invoke them or when the AI judges them relevant.
A&A perspective
A&A thinks company material can take the same shape. At the top of each document, write one line saying when to use it, for example "Read when drafting a quote." With those lines in the table of contents of your every-time document, AI has a clue for picking what it needs this time.
A&A perspective
Quote guidance holds decisions such as checking the scope of extra work before giving a price, or adding a surcharge for rush jobs. Reply guidance holds the points to ask back when it is unclear what work the client wants.
A&A perspective
In A&A's own use of AI, too, the rules read every time are kept short, and each job's procedure is read when that job begins.
A&A perspective
The table below sums up the three kinds named at the start of this article and when AI reads each one.
| Kind | Examples | When AI reads it |
|---|---|---|
| Fixed terms sent every time | How you set prices, what work you cover, work you do not take on | Every time |
| Guidance for each task | When to add a quote surcharge, what to ask back in a reply | Only for that task |
| Case records | One-off requests such as "we want to start next month" | Only when handling that case |
Keep one-off requests in case records, apart from your terms
A&A perspective
When a client says they want to start next month, that goes into the record for that case, not into your company terms. If it sits in the same place as your terms, "starting next month" can slip into a reply to a different client.
A&A perspective
Special promises on discounts or deadlines work the same way. A person decides whether a promise is for that case only or becomes your new standard. Only when you decide to make it standard do you rewrite the every-time document.
A&A perspective
Each record notes the date, the client, what was requested or promised, and whether it became standard. With a column for whether it became standard, AI is less likely to mistake a record for a current term, even when it reads the record.
From the sources
In the Claude Code documentation, Anthropic writes that when two instructions contradict each other, the AI may pick one of them arbitrarily. It therefore recommends reviewing instructions periodically and removing outdated or conflicting ones.
A&A perspective
On the day you change your prices, remove the old prices from the every-time document. On the same day, update the quote guidance that uses them. Leave past records as they were, without rewriting them at the new prices.
A&A perspective
Information that changes daily, such as stock or availability, is not copied into the document. Having AI read the original spreadsheet each time keeps old numbers from lingering.
A&A perspective
Also decide who rewrites each term, one person per term: the owner for prices, the person running operations for scope, and so on. AI and contractors may read the terms, but they do not rewrite them.
Test your setup on replies with and without a price
A&A perspective
When you change how you split the information, have AI draft both a reply that should state a price and a reply that should ask a question instead of giving one. Testing only one kind will not show whether the split worked.
From the sources
In an article on how to build evaluations for AI, Anthropic recommends testing both cases where a behavior should occur and cases where it should not. Testing only one side, it explains, leads to improvement on only one side.
From the sources
As an example, it writes that if you only test whether the AI searches when it should, you may end up with an AI that searches for almost everything. Anthropic says it built both questions that need a search and questions that can be answered without one.
From the sources
It also writes that it is often better to grade what the AI produced than the path it took to get there.
A&A perspective
For an inquiry asking the price of standard work, the right draft states the amount from the every-time document. For a request where the work itself is unclear, the right draft asks what the work is and gives no amount.
A&A perspective
Also check that another client's "starting next month" has not crept into the draft. What you check is the draft itself, not which documents AI opened.
A&A perspective
Test inquiries are easiest to gather from replies you have corrected yourself in the past. The reason you corrected each one becomes the standard for the right answer.
A&A perspective
That said, for a company whose material fits in a few pages, or for a one-off job, splitting may cost more effort than it saves. Even then, testing on both kinds of reply is still useful.
Summary
A&A perspective
To stop explaining the same things to AI every time, split your company information into three kinds. Send every time only the terms that do not change, such as prices and scope. Keep quote and reply guidance by task, and have AI read it only for that task.
A&A perspective
Keep one-off requests such as starting next month in case records, apart from your company terms. When you change the split, check it on both a reply that states a price and a reply that does not.
Pasting all your company material into every request can mix old prices or another client's wishes into the draft. Split it into fixed terms you send every time, guidance for each task, and case records, and AI reads only what this job needs.
FAQ
Can I put all my company information in CLAUDE.md?
Keep it to the terms every session needs. The Claude Code documentation sets a target of under 200 lines per file and recommends moving procedures that matter for only part of the work into a separate mechanism.
Do I need a dedicated search system before I can start?
No. You can start with the documents and spreadsheets already in your shared folders. What matters is being able to tell which terms are current, which job each document is read for, and who rewrites it.
Should I stop giving AI past conversations?
You can keep giving them. Pass them as case records, shown apart from your current company terms, so that AI does not treat a request in a record as a current term.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Effective context engineering for AI agents
Anthropic · 2025-09-29
Accessed 2026-10-22 - Equipping agents for the real world with Agent Skills
Anthropic · 2025-10-16
Accessed 2026-10-22 - How Claude remembers your project
Anthropic (Claude Code Docs) · Publication date not stated
Accessed 2026-10-22 - Demystifying evals for AI agents
Anthropic · 2026-01-09
Accessed 2026-10-22