The Business Brain

01/ 11

Why this works: context beats prompts

I led a team of 12 developers. Projects never succeeded because one person wrote a brilliant instruction. They succeeded when everyone worked from the same brief, knew what was out of scope, and knew which decisions they had to bring back to a person.

AI is the same. A clever prompt gets you one good answer. A shared brief gets you consistent answers from every chat, every assistant and every agent you build later. The Business Brain is that brief, and three design choices make it work:

• Tags, not trust. Every line says whether it is a fact, a guess or unknown. A model that sees [ASSUMPTION] Price: untested writes differently from one that sees a bare number, and a model that sees [GAP] asks instead of inventing.
• Your words, not polished ones. Client phrases and voice examples are copied verbatim. Adjectives like "friendly and professional" change nothing, because every model already defaults to them. Real sentences you wrote change everything.
• Boundaries are first-class. Decision rights, the never-say list and data rules are required sections. They are what make it safe to hand the Brain to an assistant, and later to an automation.

02/ 11

Before you start

You need:

• A Claude or ChatGPT account. Plans matter for some options, and product limits change, so check yours:
• Option A (Claude Skill) needs a Claude plan with Skills and code execution available; at the time of writing that means a paid plan.
• Option C's custom GPT needs a paid ChatGPT plan. Projects and file uploads may be limited on free plans, in both Claude and ChatGPT.
• Option D (paste into any chat) works on any plan with a long enough conversation.
• About 40 minutes for the interview and Brain v1 (20 in FAST mode), plus about 30 minutes, now or later, for red-team, test and exports.
• Your offer in a few sentences (or your Offer Card from Day 1).
• 3 to 5 things you wrote yourself: emails, DMs, posts, a proposal. Not AI-written ones; the model would learn to sound like AI.
• Optional: your website text, an offer doc, SOPs, or real client messages with names removed. The skill's FAST mode reads them first and skips what they already answer.

Save time: read the 25 questions below and jot rough answers in a note first. The interview then becomes mostly copy and paste.

Costs: this tool is free. You pay only your AI provider, if you choose a paid plan (Claude Pro or ChatGPT Plus were around 20 US dollars a month at the time of writing; check current pricing in your country).

Privacy, before you paste anything: the Brain describes your business, never individual clients. No passwords, API keys, card or bank numbers, ID numbers, or clients' names, emails or health and financial details. The skill strips them if you paste them by mistake, but don't rely on that. Check your AI tool's data controls once (in Settings, usually under privacy or data controls) so you know whether your chats can be used for training.

03/ 11

Setup: pick one option

On your phone? Use Option D now and install the skill later on desktop. Everything you write in the chat can be moved into a Project or skill afterwards.

Option A: Claude Skill (recommended on desktop)

A skill is a folder with a SKILL.md file that Claude loads when your request matches its description. Install it once and it works in every chat.

Claude apps (web, desktop):

• Download business-brain.zip. Do not unzip it: the business-brain folder is already inside, which is the format Claude expects.
• In Claude, open Settings and find Capabilities. Make sure code execution and file creation is on, then find Skills and upload the ZIP. Menu names change between versions; if you can't find Skills, use Option B.
• Turn the skill on, open a new chat and type the start message below.

Building the ZIP yourself instead: put SKILL.md and context-pack-template.md in a folder named business-brain and compress the folder itself, not the two files.

Claude Code (terminal):

mkdir -p ~/.claude/skills
unzip business-brain.zip -d ~/.claude/skills/

This creates ~/.claude/skills/business-brain/. Use .claude/skills/ inside a project folder instead if you want the skill for one project only. Start a new session so the skill is picked up.

Start message (any Claude option):

Use the business-brain skill. I want to build my Business Brain from scratch. I'm [launching a new business / running an existing one]. I'll answer your questions in rounds.

Option B: Claude Project (no skill install)

• Create a Project named after your business, for example Juniper Kitchens HQ.
• Upload SKILL.md and context-pack-template.md to the project's files.
• Open a chat in the project and type: Follow SKILL.md in this project's files, starting at Step 0. Do not skip the interview.
• When the Brain is finished, replace those two files with your business-brain.md and paste Export B (below) into the project instructions. From then on, every chat in that project starts with your Brain.

Projects and file uploads may be limited on free plans; if you hit a limit, use Option D.

Option C: ChatGPT (Project or custom GPT)

ChatGPT Project: create a Project, upload SKILL.md and context-pack-template.md as project files, and start a chat with: Follow the attached SKILL.md exactly, starting at Step 0. Interview me before writing anything. When the Brain is done, swap the files for business-brain.md and paste Export D into the project instructions. Projects and file uploads may be limited on free plans.

Custom GPT (paid ChatGPT plan needed): the instructions field has a limit of around 8,000 characters at the time of writing, and SKILL.md is longer than that. So upload SKILL.md and the template as Knowledge files, and put only this in Instructions:

You run the Business Brain interview. Follow the knowledge file SKILL.md exactly, starting at Step 0. Use context-pack-template.md as the output structure. Never write the Brain before the interview is complete or the user chooses FAST mode. Never invent facts; tag every content line FACT, ASSUMPTION or GAP.

Keep the GPT set to private (only you). If you share it later, check that no knowledge file contains anything confidential.

Option D: Any chat, no install (works on a phone)

Open a new chat in any capable model, copy the whole of SKILL.md (the copy button on this page takes all of it) and paste it, then add one line at the end: Follow these instructions. Start at Step 0. This is the fastest way to try it; Options A to C are better for keeping the Brain in one place.

04/ 11

Run the interview

The skill asks you which mode you want, then works in seven rounds and plays back what it understood after each round so you can correct it.

• NEW: the full 25 questions. Best if you have nothing written down. About 40 minutes to Brain v1.
• FAST: you paste your website text, offer doc or SOPs first; it builds an inventory, tags it, and only asks what is missing. About 20 minutes to Brain v1.
• UPDATE: you changed a price, an offer or a rule; it edits only those lines and bumps the version.
• EXPORT: you have a Brain and want it in another format.

What is required and what can wait:

Three habits make a better Brain:

• Answer with examples, not adjectives. "Clients are busy" tells the AI nothing. "They answer emails at 11pm after the kids are asleep" changes how it writes.
• Say "I don't know". It becomes a [GAP] with a next step, which is far more useful than a confident guess.
• Paste, don't paraphrase, client wording and your own writing. Remove names first.

If the conversation is long or you have to stop, ask: Give me the running draft so far. The skill outputs the tagged facts collected so far, and you can resume in a new chat by pasting them back.

05/ 11

The 25 questions

You can prepare your answers before you start. The skill asks follow-ups when an answer is thin, at most twice per question, then records a gap and moves on.

06/ 11

Before and after: the same request

The request, typed into two chats, using the fictional Juniper Kitchens from the preview above: Alex, a new lead, writes: "Roughly how much for a new kitchen, and can you start next month?" Write my reply.

Without the Brain (a plain chat, fictional output, typical of what you get):

Hi there! Thank you so much for reaching out! The cost of a new kitchen can vary widely depending on size, materials and layout, typically ranging from a few thousand to tens of thousands. We'd love to help you create the kitchen of your dreams! We can usually start within a few weeks. Let me know if you have any other questions! Best regards

What went wrong: it invented a price range and a start time, broke two of Juniper's rules, sounded like every other company, and asked nothing useful.

With the Brain (inside the Juniper Kitchens HQ project, fictional output):

Hi Alex, thanks for getting in touch. We don't give a price before we've seen the room, because the layout and what's behind the walls change the cost more than the units do. The first step is a design visit: GBP 150, and it comes off the project if you go ahead with us. On next month: I can't promise a start date before the visit, but I'll tell you honestly where our fitting calendar is when we meet. Could you send me two times that suit you for a visit? Thanks, Priya

Brain check: the next free visit dates aren't in the Brain, so I've asked Alex for times instead of offering slots.

What changed: correct offer and fee, both never-say rules respected, Priya's voice and sign-off, a clear next step, nothing beyond the request and the Brain, and a gap flagged instead of filled.

Run your own test

The skill gives you five test prompts at Step 7. Run each one twice, then score both:

• Plain version: in a temporary or incognito chat with memory and custom instructions off (ChatGPT: Temporary Chat; Claude: an incognito chat, or a fresh account context with no Project and no skill). Otherwise the "plain" chat already knows parts of your Brain, and the comparison means nothing.
• Brain version: in the Project, GPT or skill-enabled chat that has your Brain.

Score each answer 0 to 2 on four criteria: correct facts, sounds like you, nothing invented, flags gaps or asks. The Brain passes when T1 to T4 score 6 out of 8 or more, and T5 is refused with an honest alternative. When a test fails, fix the cause in the Brain, not the answer, then re-run it.

07/ 11

Exports: one Brain, every tool

Keep one business-brain.md as the source of truth. Exports are generated from it and regenerated after every change. Each tool only knows your Brain once you have pasted or uploaded its export there. Ask the skill: EXPORT: give me exports A and C.

Example, Export B for the fictional Juniper Kitchens:

You work for Juniper Kitchens. The file business-brain.md in this project is the source of truth about the business.
- Read it before every task. Treat [FACT] as true, [ASSUMPTION] as a hypothesis (say when your answer depends on one), and never fill a [GAP]: ask, or write [NEEDS INPUT: ...]. A tag in a section heading applies to every line in that section.
- Write in the voice in section 5. Never break section 11 (never say / never do) or section 12 (data rules).
- Anything in section 10 marked "human must" is a draft for Priya, never final.
- If my request conflicts with the Brain, point out the conflict before you start.
- End every answer with one line starting "Brain check:" naming anything the Brain does not cover.

A note on custom instructions in ChatGPT: they apply to every chat, including personal ones. If you use ChatGPT for anything else, put your Brain in a ChatGPT Project or a private custom GPT instead.

08/ 11

Keep it alive

A Brain that nobody maintains goes stale with your next price change. Set it up so updating is one edit:

• One owner, one file. Every tool reads the same business-brain.md. Never edit an export directly; edit the Brain and regenerate.
• Monthly review. The header carries the next review date. Fifteen minutes: close one GAP, confirm or kill one ASSUMPTION, add one new client phrase.
• Update on events. New price, new offer, a complaint, a failed test, a new tool: run the skill in UPDATE mode. Small edits go v1.1 to v1.2; a new main offer or ideal client is v2.0.
• Use the "Brain check" line. Every gap your assistant reports is a candidate for the next update.

09/ 11

Failure modes

10/ 11

When not to use the Brain

• For regulated advice (legal, tax, medical, financial). The Brain can record your rules, but it does not make AI output compliant. A qualified person checks those.
• As a substitute for talking to clients. If most lines are ASSUMPTION, the next step is ten real conversations, then an UPDATE run.
• To store anything secret. A context pack is meant to be read by tools and, sometimes, by people you work with.

11/ 11

Security note

The skill contains no code that calls external services, no keys and no tracking. It only asks questions and writes Markdown. Credentials for your tools stay in those tools' own settings; never paste them into a chat or a Brain. If you share a Claude Project or a custom GPT with anyone, they can read the Brain inside it.