Install the Shop Context Pack: company.md, customer.md, rules.md (and When Not to Build)

Formerly: The Two Files That Make Any AI Useful (and When Not to Build). Start here: this is playbook 1 of the install sequence.

You're the owner. You watched a chatbot write something that sounded like it was for someone else. That's not because you picked the wrong model. It's because the model has never read how you make money or who actually pays you. Every session starts from zero. It does not know you don't discount crown fees. It does not know you don't bid apartments. Until that lives in a file the tool can see every time, you are paying for autocomplete with manners.

This playbook installs three plain files — company.md, customer.md and rules.md — with a short never-list so the helper cannot invent a package or email a patient alone. You get a filled HVAC example, a blank dental pack, and a blank law pack. You update them on Monday. You default to buying tools. You build custom software only after a cheap loop already works. There is a one-page "do not build" checklist at the end. Use it when a vendor (or your own boredom) wants a portal with your logo.

Effectiveness: the draft talks about the thing you actually sell. Efficiency: you stop re-explaining the shop every morning. Quality: it sounds like you, and it refuses the claims that would get you a bad review or a complaint.

Who this is for

Dental, HVAC, law, agencies, multi-location retail. You have already pasted the same "we're a family business" paragraph into ChatGPT twelve times. You may also have a wishlist titled "custom AI." That list waits until these files exist and a small manual loop has been used for real.

What done looks like

• company.md and customer.md in a folder you own. HVAC version filled. Dental and law versions filled from the blanks, if that's you.
• rules.md: never-say list, approval matrix, Green/Yellow/Red data classes, tone, escalation contacts and “when unsure, draft and ask.”
• Monday: 20 minutes, files updated from last week's corrections.
• "Do not build" checklist used on the next custom quote. Default is buy/configure.

Time and tools

Three hours once. Twenty minutes Mondays. A folder. A shop-owned business AI workspace approved for your data. No developer. No "brain" that swallows ten years of email.

Prerequisites

Owner writes company.md. Not the marketing intern. Office manager (or whoever hears the phone) writes customer.md from last week's real phrases. If you would not want a line in a customer email, it does not go in the file. Treat the pack like it could leak.

Safety & data

Use a business AI workspace the shop owns, with access limited to the people doing this job. A paid plan alone is not permission to upload records. Strip names, phone numbers, addresses, dates of birth, account numbers, staff HR details and privileged material before pasting. Removing a name alone may not de-identify a record. Use synthetic examples when the remaining facts could identify someone.

For PHI, your responsible privacy lead must verify the exact product, features and integrations are covered by a signed BAA and your required safeguards before use. Otherwise keep the data out. Customer records never go into a free or personal chatbot account. Keep legal matters and payment-card data in their approved systems. AI drafts; the named workflow owner approves, and the named sender sends.

Set the shared boundary with the customer-data Field Note and the one-page AI Use Policy.

Step 1 — Write company.md (how the shop makes money)

Why. About pages lie. This is the operating truth: what pays the bills, who you won't serve, how cash arrives, what you refuse to be. Be rude on paper so the model isn't rude in public.

Clicks. New doc company.md. Owner fills it. Use the HVAC filled example as the shape. Do not paste awards, founding myths, or "excellence."

HVAC filled example (hypothetical shop — rewrite every bracket)

# company.md — Northside Heating & Air (HYPOTHETICAL)

## What we sell (this quarter's real engine first)
1. Service diagnostics and repairs (the daily engine)
2. Replacements / change-outs (where the year is made)
3. Maintenance plans (route density, not a hobby)
We do not sell "comfort as a lifestyle." We sell heat, cool, and a straight number.

## Who we serve
Single-family and small residential, north of the river, zips we actually run on the board.
We do not do rooftops over two stories, we do not bid apartment turnovers, we do not chase teaser-price emergency ads.

## How cash arrives
Card or check on the truck for diagnostics/repairs. Replacements often financed; we do not pretend a verbal "just do it" on a change-out is a job. Maintenance is billed on the plan calendar. Net-30 commercial is not our default and we do not want it to be.

## What we believe that changes a decision
Show the number before we start. Don't scare-sell a condemned unit if a repair is honest. Don't bid against a number we haven't seen. Don't promise same-day install in July unless the board says we can.

## What we refuse to be
Not the cheapest. Not a call-center voice. Not a company that leaves "misc. materials" on an estimate. Owner owns exceptions over $X. Office manager owns the calendar. Dispatcher owns the pin.

## Tools we already pay for
[Jobber / ServiceTitan / Google Voice / Gmail]. Default: use these. Do not recommend a new platform until a loop works in these.

## Claims we never make
"Tech there in 20 minutes." "Same-day install." "We'll beat any quote." Invented reviews. Prices the price book doesn't have.

## Changelog
[date] | [owner name] | [facts/claims changed and approval]

Definition of done. One to two pages. A stranger could not confuse you with the shop that will bid anything. Tools you already pay for are named.

Step 2 — Write customer.md (how they buy, not a persona poster)

Why. "Suburban Sarah, 42, values convenience" is useless. Write the words the front desk already hears.

Clicks. Office manager writes de-identified patterns from last week’s calls. Keep recordings and customer records in their approved system. Use synthetic wording when a phrase could identify someone. Do not invent income brackets.

HVAC filled example (hypothetical)

# customer.md — who actually pays us (HYPOTHETICAL)

## Trigger
It's 82 in the house and the unit is blowing warm. Or the system is 15 years old and a repair quote just landed. Or the spouse said "just call someone."

## What they will admit
Price. How long until someone is in the driveway. Whether we are licensed. Whether we make a mess.

## What they often won't admit
They already called two other companies. They are afraid of being upsold. They don't know if they are the owner or the tenant. They need to talk to a spouse before a replacement.

## Objections in their language
"That's more than my neighbor paid."
"Can you just tell me on the phone?"
"Do I have to be here?"
"Can you come tonight?"

## Questions they always ask
How soon can you get here? Do I need a new system? What if I do nothing until spring? Is financing real or a trick? Will you match the online quote?

## What actually decides yes
Who can be in the driveway today (repair). A number they can repeat to a spouse, plus a warranty they can say out loud (replacement). A human who called back. Not "our brand."

## Fears
Being upsold. A lien. Missing pickup because we ran long. A review war if it goes badly. Looking foolish on a $14,000 decision.

## Do not put in a text
Psych profiles. Guessed income. Clever nicknames. Anything creepy if a customer saw it. Keep individual preferences in the approved customer system, not this shared context pack.

## Changelog
[date] | [office manager name] | [pattern changed and owner approval]

Definition of done. You can hear your phone in the file. Objections are sentences they actually said.

Step 3 — Blank dental pack and blank law pack (fill these, don't borrow HVAC)

Why. A dental office and a PI firm should not share a company file. Copy the headings. Write your truth.

Blank dental

# company.md — [PRACTICE NAME]

## What we sell (order that pays)
Hygiene / recare:
Restorative we actually want more of:
High-ticket treatment we want more of:
What we refer out:

## Who we serve
Geography:
Families / adults / both:
We are not: [fill with truthful service boundaries, without demeaning patients or other practices]

## How cash arrives
Insurance billed later vs card-on-file vs same-day patient portion:
What we do when a statement becomes a fight:

## Beliefs that change a decision
We explain out-of-pocket without shame.
We do not: [scare-sell / promise insurance covers everything / diagnose by email]

## Refuse to be
Claims we never make: insurance covers everything; it won't hurt; same-day veneers we saw on a reel; fabricated Google reviews.

## Who owns what
Office manager owns the calendar.
Owner owns exceptions over $___ and anything clinical in public.
Front desk owns: [neutral review invitation policy / consent checks / manual sends]. No satisfaction-based filtering.

## Tools we already pay for
[Open Dental / Dentrix / Weave / ...]. Default: use these.

# customer.md — [PRACTICE NAME]

## Trigger
(I haven't been in three years and I'm nervous / tooth pain / kid / transfer)

## Admit / won't admit
Admit:
Won't: (fear, embarrassment, they already called another office)

## Objections
"Will insurance cover it?"
"Do I have to come in?"
[add yours]

## Always-ask questions
Will it hurt? How soon? What do I bring? How long will I be numb / out of work?

## What decides yes
Who explained the out-of-pocket without shame. Who had a morning slot. Who didn't make them sit 40 minutes.

## Fears
Pain. Being upsold. A lecture. Missing school pickup. A bad review they feel stuck writing.

## Never in a patient text
Diagnosis, "your x-ray looks fine," clinical advice, a promised wait we can't keep, creepy persona language.

Blank law

# company.md — [FIRM NAME]

## Matters we take (this quarter)
We take:
We refer out:
We do not take: (every angry "I want to sue")

## Who we serve
Geography / venue:
How they find us:

## How cash arrives
Consult fee or not:
Retainer:
We never: promise a result to get a card.

## Beliefs that change a decision
We don't promise court outcomes.
Conflict check is human. Always.
Incomplete intake does not get a consult checkmark.

## Refuse to be
TV-lawyer voice. "We will win." Advice by email. Auto-filing. Public case recaps on Google.

## Who owns what
Intake coordinator owns the queue.
Attorney owns: conflict, advice, anything to the other side, anything a court will read.

## Tools we already pay for
[Clio / PracticePanther / Gmail / ...]. Default: use these.

# customer.md — [FIRM NAME]

## Trigger
(I got a letter and I don't know if it's serious / accident / landlord / contract)

## Admit / won't admit
Admit: they want to know if it's worth a consult.
Won't: they already talked to two firms; they're embarrassed; they want a fight more than a result.

## Objections
"Can you just tell me on the phone?"
"How can it cost that if it's simple?"
"Will I have to go to court?"

## Always-ask questions
How soon can we talk? Do I have to pay today? What do I bring? What have I already filed?

## What decides yes
Who called back before lunch. Who explained the next step without a speech. Complete intake.

## Fears
Looking foolish. A lien. Talking to the other party. A public Google description of their matter.

## Never in a client email from a bot
Legal advice, outcome claims, "you're clear" on conflicts, anything to opposing counsel, recap that includes advice.

Agencies: one pack per client or the helper will mix voices. Multi-location retail: one company file, a short overlay per store (hours, SKUs that store never carries, manager name), one customer file if the buyer is the same person.

Definition of done. Your vertical's blanks are filled. HVAC owners still have Step 1–2 files, not these.

Step 4 — Write rules.md (the house rules for every tool)

This third file turns values into permissions. A file is guidance, not a security control: use the tool's actual permissions to prevent sends, payments and record access your workflow does not authorize. Name the people. “The team” cannot approve an exception.

Filled HVAC example (Hypothetical)

# rules.md — Northside Heating & Air (HYPOTHETICAL)
OWNER: Shop owner. BACKUP: Office manager. Replace roles with actual names.

NEVER-SAY LIST
No invented prices, discounts, warranties, reviews or arrival times.
No "tech there in 20 minutes," "we beat every quote" or unverified diagnosis.
No private customer details in public replies. No pretending to be a customer.

APPROVAL MATRIX
Customer email/text/review: office manager reviews; named dispatcher or
front-desk lead sends. Owner reviews complaints and pricing exceptions.
Money: owner verifies against the approved price book; AI never charges,
refunds, cancels a subscription or changes payment details.
Schedule: dispatcher checks the real board and approves; AI drafts options,
never books, moves or cancels work on its own.
Technical safety: licensed responsible person decides; AI does not diagnose.

DATA CLASSES
Green: approved public services, opening hours and website copy.
Yellow: de-identified internal process patterns, timings and synthetic examples.
Red: personal identifiers, PHI, privileged legal material, card data, secrets,
and staff HR details. Keep out of this context pack and ordinary AI drafting.
Any exception needs a separately approved exact product/workflow, necessary
agreements (including a signed BAA where HIPAA applies), permissions and safeguards.
A business subscription alone does not make Red data acceptable.

TONE
Plain, respectful, short. No pressure, fake urgency, diagnoses or legal advice.

ESCALATION
Operational uncertainty: [office manager's approved internal contact].
Price/complaint: [owner's approved internal contact].
Data/security concern: [responsible lead's approved internal contact].
Emergency: follow the shop's existing emergency procedure; do not wait for AI.

WHEN UNSURE: draft and ask. Do not act, send, post, pay or change a calendar.

CHANGELOG
[date] | [owner name] | Initial approved rules and permissions checked.

Blank dental rules

# rules.md — [PRACTICE]
Owner / backup: [names]
Never say: [unverified diagnosis, treatment promise, insurance coverage guarantee,
patient relationship in public, clinical detail in reminders; add your rules]
Customer messages: [named reviewer] approves; [named front-desk sender] sends.
Clinical content: [licensed responsible clinician] decides; AI never diagnoses.
Money / exceptions: [name] verifies; AI never charges, discounts or refunds.
Schedule changes: [name] checks actual capacity and approves before human action.
Green: [approved public facts]
Yellow: [de-identified process facts or synthetic examples]
Red: PHI, identifiers, card data, staff HR details, secrets; no Red data in pack.
Separate approved Red-data workflow, if any: [exact product/features, responsible
privacy lead, agreement/BAA coverage and safeguards — blank means prohibited].
Tone: [plain, respectful; no pressure or fear]
Escalation contacts: [clinical / privacy / owner, approved internal contacts]
When unsure: draft and ask. No sends, records access or actions on your own.
Changelog: [date] | [name] | [what changed and who approved]

Blank law rules

# rules.md — [FIRM]
Owner / backup: [names]
Never say: [outcome guarantee, "conflict clear," legal conclusion, matter detail
in public; add your rules]
Client/prospect messages: [attorney reviewer] approves; [named sender] sends.
Legal advice, conflicts, deadlines, filing and signing: [qualified attorney]
decides. AI may draft a checklist; it never files or declares a conflict cleared.
Money / exceptions: [authorized person] verifies; AI never pays or changes details.
Schedule changes: [name] checks the actual calendar and approves human action.
Green: [approved public facts]
Yellow: [de-identified operations and synthetic examples]
Red: privileged/confidential matter material, personal identifiers, card data,
HR details and secrets; no Red data in pack.
Separate approved restricted-data workflow, if any: [attorney-approved exact
tool, terms, permissions and safeguards — blank means prohibited].
Tone: [plain, no promises or television-lawyer language]
Escalation contacts: [attorney / privacy-security / owner]
When unsure: draft and ask. No sends, external forwards, payments or filings.
Changelog: [date] | [name] | [what changed and who approved]

Definition of done. Rules name an approver and sender for messages, an approver for money and schedule changes, and a real escalation contact. Test a synthetic request for an unauthorized discount: it should draft a question and stop.

Step 5 — Install the pack everywhere you use AI; keep it true on Monday

Why. Files compound. Chat windows don't. A Monday ritual is how last week's "never say delighted / never discount / always ask gate code" gets into this week's draft.

Clicks.

1. Keep one source of truth in the shop’s shared drive: shop-context/company.md, customer.md, rules.md. Owner controls edits. Optional voice.md uses three synthetic or de-identified good/bad examples; never attach raw customer mail or records.
2. Install approved copies in Claude or ChatGPT Project instructions, a Gemini Gem, or workspace/agent instructions where your product and plan support them. Confirm the actual scope and settings; do not assume a workspace default reaches every chat or connected app. Instructions: “Read company.md, customer.md and rules.md first. When unsure, draft and ask.”
3. For coding agents working on your shop’s site, put the relevant non-secret rules in the instruction file that the agent actually reads, such as AGENTS.md or CLAUDE.md. Verify the documented filename, scope and load behavior; do not assume a lowercase agents.md file is recognized. These copies never contain credentials. Keep customer messaging instructions separate from permission to deploy or change accounts.
4. Record every installed copy: tool, location, owner, source version and last sync. Run one synthetic draft in each and check a never-list violation is caught. Start the next permitted job with “read the files first, then draft.” A named human checks and sends. If it is generic, improve the context.
5. Add a changelog to the bottom of each source file: date | who | what changed | who approved. After an approved change, sync every listed copy and repeat the sample test.
6. Monday 20 minutes (print this):

MONDAY CONTEXT — 20 MIN  —  owner + office manager
- Add last week’s three de-identified corrections to the right file (voice, claims, objection)
- Add a dated changelog line to each changed file; owner approves
- Sync the approved version to the installed-copy register; verify one sample
- Delete anything we wouldn't want in a customer email
- Confirm tools-we-already-pay-for is still true (did we actually start using the texting add-on?)
- Do not add a fourth file unless a loop proved it needs one (warranty language, fee schedule, store overlay)
- Run ONE sample draft. If it still sounds like a billboard, fix company.md or customer.md, don't buy a new model

I would not fine-tune anything. I would not dump ten years of mail into a "brain." Three small files are enough to start. Instructions guide behavior; enforce access and actions with actual permissions too.

Definition of done. The project has the files. Monday is on the calendar. One live workflow is using them.

Step 6 — Default buy. Custom only after a cheap loop. Do-not-build checklist.

Why. Most shops need a workflow, not a product. An approved tool you already pay for plus a one-page SOP will beat a science project nobody opens on Tuesday. Custom starts to make sense only when the bottleneck is clicking, not thinking — and only where the shop's way of working is the actual advantage.

Clicks. Take the next vendor quote or the next "we should build an agent" idea. Run this page. If it fails, you don't build. You buy, or you write an SOP, or you do nothing.

One-page "do not build" checklist

DO NOT BUILD IF ANY OF THESE ARE TRUE
[ ] We do not have company.md + customer.md + rules.md in use on a real weekly job
[ ] Existing approved tools and a one-page SOP could handle the useful core in two weeks
[ ] The PMS / dispatch / practice / POS we already pay for already does this and we never turned it on
[ ] We want AI because the last software install failed and we'd like a prettier failure
[ ] Nobody is named as the person who will run this every Tuesday
[ ] There is no human checkpoint before email / file / charge / diagnose / post a review
[ ] We cannot name the door: money, time, or customer experience — and how we'll notice in 30 days
[ ] The request is auto-file, auto-diagnose, auto-email a patient, auto-publish a legal letter, or fake reviews
[ ] Data lives in four spreadsheets and the owner's head
[ ] We have not run the cheap loop by hand long enough to know if people will use it

BUILD A SMALL HELPER ONLY IF ALL OF THESE ARE TRUE
[ ] Cheap loop already used for real (weeks, not a demo)
[ ] Bottleneck is now clicking, not thinking
[ ] Our way of working is the advantage a generic tool keeps flattening
[ ] Data lives in one place we can actually connect, or we will clean it
[ ] Human checkpoint is designed in
[ ] Scoreboard named
[ ] Someone owns it after month one (the owner or a paid operator — not "the team")

EVEN THEN: smallest spine. A button that parks a draft in the doc the office manager already uses.
NOT a platform with users, permissions, a mobile app, and a roadmap. You don't want to run that company.

Buy/configure examples (do these first). HVAC: dispatch board + estimate software + approved AI workspace + SOP for first-draft estimates. Dental: recall list you already have + recare script + human send. Law: better intake form + transcription + checklist; not an "AI associate" that files. Agency: prompt + folder + Friday status note. Retail: POS + shared sheet + store-level text.

Hypothetical, labeled. Office manager spends 45 minutes turning tech notes into an estimate email. SOP: paste notes, attach price rules, never discount, permit line on change-outs. Model drafts in four minutes, she edits five. That's the loop. Custom might later read photos from a folder you already drop files into. Custom is not an agent that books the install and orders parts.

Definition of done. The checklist lives in the same folder as company.md. The next custom quote got this page before it got a yes.

Hypothetical worked example

Hypothetical. Not a client.

Four-location idea, but they start with one HVAC shop. Owner writes company.md on a Thursday night (the filled example, rewritten). Office manager writes customer.md from a week of "can you just tell me on the phone." They add rules.md. Claude Project gets the folder. First job: estimate follow-up drafts only. Week two the model offers 10% off. They don't do that. Monday they add "never discount" to company.md. Week three the draft asks for the gate code without being told — because Monday put it in.

A vendor pitches a custom "AI dispatcher OS." They run the do-not-build checklist. Cheap loop isn't even a month old. PMS already has dispatch. No Tuesday owner named for a new app. They don't build. They buy nothing. They keep the SOP.

Dental branch of the same hypothetical: recare texts from customer.md language, never a diagnosis. Law branch: intake summaries, never an email to the other party. Same three-file structure, different refusals.

Operator Standard check

Trace: Write the real services, buying patterns and hard boundaries without putting customer records in the pack.

Choose: Approve the never-list, data classes, named approvers and small manual workflow before buying or building.

Build: Keep one source of truth, install verified copies and use real permissions to constrain each tool.

Prove: Test one draft every Monday, track recurring errors and sync corrections before expanding.

Failure modes

Marketing intern writes company.md. You get the About page again. Owner writes it.

Persona posters. Age brackets you guessed. The model will write like a textbook.

Dumping ten years of email. You scaled gossip and PHI risk. Two pages beat a junk brain.

No Monday ritual. The files rot. The helper offers 10% off forever.

Building first. Portal with your logo while missed calls still hit voicemail. Effectiveness failed.

Allowing auto-send because “the files are good now.” Do not enable it. A named human approves and sends every customer message; files do not grant permission.

For the decision behind this workflow, read the Field Note on company and customer context.

FAQ

Isn't this just a brand book? No. Brand books say excellence. These files say what you refuse, how cash arrives, and the sentence the front desk actually hears.

Do I need offer.md and voice.md too? Need: company, customer, rules. Voice examples help review replies and texts. Offer/claims-to-avoid belongs in company.md until it outgrows it. Don't build a wiki.

Can I put patient or client records in the folder? No. This shared pack contains approved public facts, synthetic examples and de-identified patterns. Patient charts, matter files and HR/payment details stay out. A separate workflow involving restricted records needs the responsible lead’s approval, appropriate agreements and safeguards for the exact tools involved.

When do we actually build software? After the cheap loop works, the bottleneck is clicking, and the checklist's second half is all true. Even then, a small helper. Not a product company.

If it goes wrong (kill switch)

The owner or workspace administrator can pause the pack: stop pasting, remove the suspect instruction copy or disable the affected assistant, and restore the last approved version from the shared drive’s history. Use the prior manual SOP while checking every installed copy. If a wrong promise reached a customer, the named sender identifies the incorrect statement, provides the verified correction and next step after owner approval, and records the incident. Privacy or legal exposure goes to the qualified lead before further disclosure.

Prove it: 4-week scoreboard

Record the baseline before changing the workflow. Use actual shop counts; leave blanks until measured.

Save it as a skill

Save the final prompt and never-list in skills/shop-context-pack.md in the shop’s shared drive, or the approved project instructions your AI workspace supports. Name its owner and review date. The file is reusable guidance, not a grant of sending or payment access.

ROLE: Draft assistant for [SHOP].
READ FIRST: company.md, customer.md, rules.md from approved version [date].
PROMPT: Draft only the requested bounded task using verified facts. Mark unknown facts and conflicts. Show the named reviewer and sender.
NEVER-LIST: [paste the final rules.md never-say list here]. No invented prices, outcomes, reviews, diagnoses or arrival promises. No customer records in the pack.
DATA: Green public facts; Yellow approved de-identified patterns; Red excluded unless a separate exact workflow is explicitly approved.
ACTIONS: Never send, post, pay, file or change schedules on your own. Named human approves and sends every customer message.
WHEN UNSURE: Draft a question and ask [escalation contact].
OWNER / SOURCE VERSION / NEXT REVIEW: [fill]

Next install the AI Use Policy, then use the Three Outcomes Filter to choose work and Tool Stack Audit before adding a subscription. Read the context Field Note.

Screen-recording outline (9–12 minutes)

1. 0:00–0:40 — "Your AI is dumb because it doesn't know how you make money. Three files. Default buy. Don’t build yet."
2. 0:40–3:30 — Fill HVAC company.md live from the template. Read the refuse-to-be paragraph out loud.
3. 3:30–5:30 — customer.md from fake voicemail phrases. Kill "Suburban Sarah."
4. 5:30–7:00 — Show blank dental + blank law headings. Fill one refusal line on each.
5. 7:00–8:30 — Upload to a Claude Project. Draft an estimate follow-up. Show it trying a discount; add never-discount; re-draft.
6. 8:30–10:30 — Close: "Let’s talk about your business. Start with a free introductory conversation. The $1,500 AI Workflow Sprint helps your team use AI on real work; implementation is scoped separately." Direct viewers to /start.