Set up a Claude Project as your business brain

What you will build

A business brain: one Claude Project that holds your business facts, prices, policies, FAQ, tone examples and Desk rules as files, with written project instructions that tell Claude how to use them. You will prove it works with a test set of real questions and expected answers, fix what fails, and save a dated known-good copy you can restore.

The same setup works in a ChatGPT Project or a Gemini Gem. Step 9 shows the differences.

Before you start

A Claude account. Claude's help centre says Projects are available to all users, including free accounts, with a limit of five projects on the free plan. On paid plans, Claude can switch to a retrieval mode that expands how much knowledge a project can hold. See What are projects?
Your business material in any form: price list, service descriptions, policies, common questions, a few good replies you have sent, and any Desk rules you already use.
15 real customer questions for the test set, anonymised.
A folder on your computer for the knowledge files and test results.
About 60 to 90 minutes.

How it works

Claude's help centre describes a Project as a self-contained workspace with its own chat history and knowledge base. Project instructions apply to every chat inside it, and context is not shared across chats unless the information is in the knowledge base. That is what you want: the files are the memory, not an old chat.

This guide treats the Project as a Desk. The files are your facts. The instructions are your rules, including stop rules and human-owned actions. The test set is your proof. The known-good copy is how you recover after a bad edit.

Decide what the Project is for, and what it is not for.
Prepare clean, single-topic knowledge files.
Create the Project and add the files.
Write the project instructions.
Build a test set with expected answers.
Run the tests in fresh chats and grade them.
Fix the files or the instructions, then re-run everything.
Save a known-good copy and name an owner.
Build the same brain in ChatGPT or Gemini, if you use them.

Step 1: Decide what the Project is for, and what it is not for

A Project that holds your whole business becomes vague. Give it one clear job, such as "answer questions about our services and draft replies and quotes from our own files".

TEXT

PROJECT PURPOSE CARD
Project name: [e.g. Customer Desk brain]
Used by: [owner, front desk]
Used for: [answer service questions, draft replies, draft quotes]
Not used for: [customer records, staff matters, bank or tax documents]
Owner (keeps files current, approves changes): [name]
Human-owned actions: send, pay, delete, publish, change accounts

Need a second brain for admin or reporting? Make a second Project.

Check

The Project has one job, written in one line.
The "not used for" list names your sensitive material.
An owner is named.

Step 2: Prepare clean, single-topic knowledge files

Good files matter more than clever instructions. Write each file in plain text or Markdown, one topic per file, with a "Last checked" date at the top. Claude also accepts formats such as PDF and DOCX, but plain text is easiest to keep clean.

A starter set:

TEXT

01-business-facts.md    what we do, areas, hours, contact, how booking works
02-price-list.md        one block per service, units, "needs owner price" rules
03-policies.md          deposit, cancellation, warranty, payment methods, refunds
04-faq.md               the 15 to 20 questions customers ask most, with our answers
05-voice-examples.md    3 to 5 real replies we liked, names removed
06-desk-rules.md        draft rules, stop rules, human-owned actions

Rules for the files:

One fact lives in one file only. If the price is in the price list, the FAQ says "see price list", not a second copy of the price.
No ranges or "about" prices. Use "needs owner price" instead.
No old promotions, no draft ideas, no chat exports.
Clear, descriptive file names, so you and Claude can tell them apart.

The resource Put your price list in a shape AI can read correctly walks through the price file in detail.

Check

Every file has a single topic and a "Last checked" date.
No fact appears in two files.
No customer or staff personal data appears anywhere.

Step 3: Create the Project and add the files

In Claude, open Projects from the left sidebar (or go to claude.ai/projects) and choose "+ New Project". Give it the name from your purpose card and a one-line description. On Team and Enterprise plans you will also choose whether the project is private or shared with your organisation. Keep it private until it passes your tests.

Then add your knowledge files from the project page using the "+" button. The steps are from Claude's help article How can I create and manage projects?. Menus may move; check your app.

Upload only the six files from Step 2. Extra files mean more chances of conflicting facts and more of your usage allowance spent on every chat.

Check

The Project is named and described clearly.
Only the prepared files are uploaded.
The project is private until testing is done.

Step 4: Write the project instructions

Project instructions apply to every chat in the project. On the project page, choose "Set project instructions", paste your instructions and choose "Save instructions".

Anthropic's prompting best practices recommend being clear and direct, explaining the reason behind an instruction, and telling Claude what to do rather than only what not to do. The template below follows those ideas.

TEXT

You are the Customer Desk for [business name], a [type of business] in [area].

SOURCES
- Answer only from the project files. They are the single source of truth.
- If the files do not answer a question, say "Not in my files" and suggest which file the owner should update. Do not fill gaps from general knowledge.
- Prices come only from 02-price-list.md. If a request matches a "needs owner price" rule, say the owner will confirm.

DRAFTS
- Customer replies are read on WhatsApp, so keep them under 60 words, in the customer's language, with one next step.
- Match the tone of 05-voice-examples.md.
- After every draft, add FACTS USED (quote the file lines) and UNKNOWN (anything not answered by the files).

STOP RULES (write OWNER REVIEW and the reason instead of a draft)
- Complaints, refunds, injuries, health, legal or money disputes.
- Discount requests or anything in "things only the owner decides".
- "Stop messaging me", "wrong number", "not interested".

HUMAN-OWNED ACTIONS
- You prepare drafts only. A person sends, pays, deletes, publishes and changes accounts.

PRIVACY
- Do not ask customers for IC numbers, bank details or other personal data. If a pasted message contains them, remind me to remove them.

The WhatsApp line explains why, which helps Claude choose the right length even for cases the rule did not name.

Check

Instructions say where answers must come from and what to say when they cannot.
Stop rules and human-owned actions are written in.
Every draft ends with FACTS USED and UNKNOWN.

Step 5: Build a test set with expected answers

Before you use the brain with customers, decide what "working" means. Anthropic's guide to defining success criteria and building evaluations recommends criteria that are specific and measurable, and tests that mirror real tasks, including edge cases. For a business brain, that means real questions with answers you already know.

Write 15 questions in a spreadsheet or text file:

TEXT

TEST SET
# | question (anonymised) | expected answer or behaviour | source file | pass rule
1-8   normal questions whose answers are in the files
9     a service you do not offer            -> "Not in my files" or a polite no from 01
10    a discount request                    -> OWNER REVIEW
11    a complaint or injury                 -> OWNER REVIEW
12    an area outside your service area     -> the rule from 01
13    a question in BM                      -> BM answer, same facts
14    a mixed BM-English question           -> answer in the same mix
15    a price with a "needs owner price" rule -> owner will confirm, no number

Write each pass rule in plain words, for example "gives the Small dog full grooming price from 02 and nothing else", not a vague "sounds right".

Check

15 questions, including the seven edge cases.
Every question has an expected answer and a source file.
Pass rules are specific enough to mark quickly.

Step 6: Run the tests in fresh chats and grade them

Open a new chat inside the Project and paste the questions, one at a time or in small groups. A fresh chat tests the files and instructions, not what you said earlier.

Grade each answer against its pass rule. Record the result:

TEXT

TEST RESULTS: [date], project version [v1]
# | pass / fail | what happened | cause (file / instruction / unclear question)
Overall state: verified (all pass) / unverified (some not yet checked) / broken (any fail)

Anthropic's guide to reducing hallucinations also suggests running the same prompt more than once and comparing the outputs, because inconsistent answers can signal a problem. Run your trickiest three questions twice in separate chats. If the answers disagree, the files or instructions are not clear enough yet.

Check

Every test ran in a chat inside the Project, started fresh.
Each result is marked pass or fail with a reason.
The three hardest questions were run twice.

Step 7: Fix the files or the instructions, then re-run everything

Map every failure to a cause and fix that cause:

The answer invented a fact → the fact is missing from the files, or the "Not in my files" rule is too weak. Add the fact, or strengthen the rule.
The answer used an old price → a second copy of the price exists somewhere. Delete it.
A stop rule was ignored → make the rule more specific, give the reason, and add an example.
The tone is wrong → replace a voice example rather than adding more adjectives.

Change one thing at a time, then re-run all 15 tests, not only the one that failed. A fix for one question can quietly break another.

Check

Every failure has a written cause and fix.
The full test set was re-run after the last change.
All 15 pass in a fresh chat.

Step 8: Save a known-good copy and name an owner

When the tests pass, save the known-good copy outside Claude:

TEXT

Business brain known-good/ [date]
  instructions.txt          exact project instructions
  01 to 06 knowledge files  the versions that passed
  test-set.txt              questions, expected answers, pass rules
  test-results-[date].txt   the passing run
  CHANGELOG.txt             what changed, when, by whom

The owner updates files when facts change, re-runs the test set after each change, and saves a new known-good copy. A monthly check is a good rhythm even if nothing changed.

A note on memory. Claude's help centre says each project has its own memory space, separate from your other chats, when memory is on, and that you can view and edit what Claude remembers in Settings under Memory. See Use Claude's chat search and memory. For a business brain, do not rely on memory for facts. If something matters, it belongs in a file.

Check

The known-good folder holds instructions, files, tests and results.
The owner and the review rhythm are written in the changelog.
Important facts live in files, not in chat memory.

Step 9: Build the same brain in ChatGPT or Gemini, if you use them

The files, instructions and test set are portable. Only the setup screens differ. Menus may move; check your app.

ChatGPT Projects: OpenAI's help centre describes Projects as keeping related chats, files and instructions together. You create a project, add files, set project instructions, and can choose between default memory and project-only memory. See Projects in ChatGPT. Paste the same instructions and upload the same six files.
Gemini Gems: Google's help page Use Gems in Gemini Apps describes creating a Gem from the Gems menu with New Gem, entering a name and instructions, and adding files under Knowledge from your device or Google Drive. Drive files are used in their most recent version. Adding Drive files needs Keep Activity on, which Google links to human review of some chats, so upload from your device if that matters to you.

Run the same 15 tests in each tool and grade them the same way. Do not assume a brain that passes in one tool passes in another.

Check

The same files and instructions are used in every tool.
The full test set was run in each tool.
Privacy settings were checked in each tool before uploading.

Worked example

This example is made up for teaching. The business and details are not real.

A pet grooming salon in Puchong kept re-explaining dog sizes, what full grooming includes, and that cats are not groomed. The owner built a private Customer Desk brain with the six files and the template instructions.

Part of 01-business-facts.md:

TEXT

Last checked: 2026-09-23
We groom dogs only. We do not groom cats or other animals.
Service area for pick-up: Puchong and Bandar Kinrara.
Booking: customer sends breed, weight and preferred day; the owner confirms the slot.
Dogs with skin wounds or recent surgery: owner decides after seeing a photo.

Test results, first run: 12 of 15 passed. Three failed:

Q9 "Can you groom my cat?" The answer said "Yes, we offer cat grooming on request." The FAQ had an old line from before the salon stopped grooming cats. Cause: a second, outdated copy of a fact. Fix: deleted the FAQ line and pointed it to 01.
Q10 "Discount if I bring two dogs?" The answer offered a "multi-dog discount". Cause: 05-voice-examples.md contained an old reply with a one-off discount. Fix: replaced that example.
Q11 "My dog has a small cut after grooming yesterday." Passed once, but on the second run it offered a free re-groom. Cause: pet injuries were not named in the stop rules. Fix: added "injuries or health concerns about a pet", with the reason "a person must look at every health issue".

After the fixes, all 15 tests passed in a fresh chat, and the three hardest passed twice. The owner saved the known-good folder. Later, for "Berapa harga grooming Shih Tzu 6kg?", the draft quoted the Small dog block from 02-price-list.md in BM, listed FACTS USED, and waited for the front desk to send it.

Common mistakes and fixes

The same fact in two files → keep each fact in one file and point to it from others.
Relying on chat memory for facts → put every important fact in a file; memory is not your source of truth.
No test set → write 15 questions with expected answers before using the brain with customers.
Fixing one failed test and moving on → re-run the full set after every change.
Old replies with one-off deals in voice examples → replace them; the AI copies what it sees.
Sharing the project before it passes → keep it private until the tests pass, then share.

Take it further

Add a Skill: a Project holds knowledge for one area. A Skill packages a procedure Claude can use in any chat. See Write your first Claude Skill for quotes or enquiry replies.
Connect live tools, carefully: once the brain is stable, you may want Claude to read Gmail or Drive. See Connect Claude to Gmail, Drive and Calendar safely first.
Desks that use the brain: The Lead Triage Desk: sort every new enquiry and draft replies you approve, Turn a customer message into a draft quote and A weekly business report you check instead of write all get simpler inside a tested Project.
Monthly habit: owner checks each file's date, updates facts, re-runs the tests, saves a new known-good copy.
Official course: Anthropic's AI Fluency for small businesses covers repeatable workflows and sensitive business information.

Quick checklist

One Project, one job, one owner.
Six single-topic files with "Last checked" dates; no personal data.
Each fact in one file only.
Instructions: files only, "Not in my files", stop rules, human-owned actions, FACTS USED.
15-question test set with expected answers and edge cases.
Tests run in fresh chats; hardest questions run twice.
Every failure fixed at its cause; full set re-run.
Known-good folder saved with instructions, files and results.