Learn My Business

This is the connected deepener, not the cold front door. Brand-new and keyless owners start with /start-here — a 60-second brain-dump that fills the profile and lands a real win with zero accounts. This skill picks up after a CRM is connected: it reads the live workspace and replaces the inferred guesses in CLAUDE.md with filled, accurate data — the real pipeline, products and prices, stages, lead sources — with any industry-specific gotchas folded in. Run it once on connect, and re-run it whenever the workspace shape changes.

Step 1 — Read the workspace shape

python ~/.claude/bos-run.py learn-my-business

Returns the real shapes: company profile + brand, every pipeline with its actual stage names, products with prices, lead sources, opportunity types, lost/won reasons, and rough record counts. It's best-effort — check warnings and _sources; for anything that came back unavailable, ask the operator one short question rather than guessing (e.g. company-profile endpoints vary, so you may need to ask "what's your business name and what do you do?").

Fallback if the script can't run: list_pipelines + list_pipeline_stages, list_products, get_company_profile, get_crm_settings.

Step 2 — Load the structure + the industry gotchas

There is one base template: templates/CLAUDE.md — read it so the file you write keeps its structure and you fill it from real data rather than inventing a new format. The template is vendor-neutral by design: it carries no TrustPager pitch or tool catalogue. The TrustPager operating context is added separately in Step 2b once you've confirmed the connection is live.

Then match the operator to a business shape in knowledge/industry-notes.md using company.industry and how their work flows (how leads arrive, how it's priced and delivered, what brings customers back):

• match the shape first — service/professional, trades/on-the-tools, product-seller/ecommerce-retail, hospitality/walk-in, clinic/appointment, courses/community/coaching, software/digital-product (industry-notes.md owns the list) — then layer the vertical specifics nested under that shape if a named industry (mortgage/finance, trades, insurance, consultant, technical services, manufacturing, ecommerce/retail, hospitality, allied health, wellness, course creator, community operator, coach, SaaS) fits.
• many businesses are a blend of two shapes (a community with a software upsell, a shop that runs workshops) — take the gotchas from each part.
• nothing fits cleanly → use the generic fallback as-is and ask one or two short questions about how their work flows and their comms style. Never tell the operator their business "doesn't fit" or is unusual — shapes are internal machinery; the profile you write sounds equally confident either way.

If the line of work is ambiguous or unavailable, ask one short question ("what would you call your line of work?") rather than guessing the shape. Pull the shape's (and any vertical's) gotchas and comms style into the file you write — but treat them as patterns to confirm, never as facts read from the workspace (see hard rules).

Step 2b — Plant the TrustPager operating context

Because this skill only runs once a TrustPager workspace is connected (Step 1 read its live shape), the owner's profile should now carry the full TrustPager operating capability — the part the day-zero template deliberately leaves out.

Read drivers/trustpager/OPERATING-CONTEXT.md (the canonical source for how to behave when the connection is live: the one-workspace/your-key framing, the Australian-first defaults, the approval-queue guardrail, credits, signals, and the practical tool list). Fold it into the owner's ./CLAUDE.md as a clearly marked TrustPager operating section, so every subsequent session carries it.

This merge follows the same no-clobber rule as Step 3: if ./CLAUDE.md already exists, never silently overwrite it — show the diff (what you're adding / changing) and ask before replacing. A hand-tuned profile may have edits worth keeping; merge the operating context in, don't clobber.

Step 3 — Write the CLAUDE.md

Fill every <<< ... >>> from the digest:

• My business — name, city/country, description from company. If industry is known, say it.
• Products / services — real product names + prices from products.
• Pipeline — the actual stages of the default pipeline, in order. If there are several pipelines, list the primary one's stages and name the others.
• Lead sources — tick the boxes that match settings.lead_sources.
• How the business is running (labelled estimates, never facts) — where the workspace yields them, compute and write into the profile's "How the business is running" block: rough win rate from won/lost counts, deals per stage, primary lead source by volume, and repeat-vs-one-off customer mix. Label every one "estimated from your workspace — confirm". These are the diagnosis inputs knowledge/business-method.md §2 needs; the profile is where they live so no future session re-asks.
• Ideal customer / tone — leave as a short prompt for the operator to confirm; you can draft a first guess from the industry, but flag it as a guess.

Then write it to ./CLAUDE.md in the operator's project folder.

• If ./CLAUDE.md does NOT exist → write it, then show a summary of what you filled.
• If it DOES exist → do NOT overwrite silently. Show the proposed file, point out what differs from the current one, and ask before replacing. (Their existing file may have hand-tuned voice/rules worth keeping — merge, don't clobber.)

Step 4 — Confirm + close gaps

Show the operator a tight summary:

✓ Wrote CLAUDE.md from your workspace:
  • Business: <name> — <industry>, <city>
  • Pipeline "Sales": New lead → Qualified → Quote sent → Negotiation → Won
  • 4 products (CRM Suite $129/mo, AI Audit $499, ...)
  • Lead sources: Website form, Meta ads, Referrals

Two things I couldn't read and guessed — please confirm:
  • Your ideal customer (I drafted a line from your industry)
  • Your tone preferences (left as a prompt)

If the profile still lacks them, ask AT MOST two of: "How many more customers could you take on right now — and what breaks first if you doubled?" (§2's capacity probe, knowledge/business-method.md); "Do your customers mostly find you locally, or could they be anywhere?" (gates the local playbook, §10.5); and note the revenue band the workspace implies as a guess to confirm (§5). Two questions maximum — this is a sync, not an interview.

This two-question ceiling is this skill's declared cap, and a labelled override of the default discovery backstop in knowledge/conversation-method.md §1, because a connected sync reads the workspace instead of interviewing the operator. At the second answer take §1's graceful exit rather than asking a third question, and run the two questions themselves under §2-§3.

End by telling them: "Claude will use this from your next session. Re-run /learn-my-business whenever your pipeline, products, or brand change."

Hard rules

• ❌ Don't overwrite an existing CLAUDE.md without showing the diff and asking.
• ❌ Don't invent products, prices, or stages — use the real ones from the workspace. If a section is unavailable, ask one question; don't fabricate.
• ❌ Don't fill the "ideal customer" / "tone" sections as if they were read from data — they're your guess; label them for confirmation.
• ✅ Keep templates/CLAUDE.md's structure intact, fold in the matched industry section's gotchas + comms style, and plant the TrustPager operating context from drivers/trustpager/OPERATING-CONTEXT.md (Step 2b).
• ✅ Industry gotchas are patterns to confirm, not workspace facts — present them as defaults the operator can correct, and let real workspace data win where they conflict.
• ✅ Use the operator's exact stage and product names, spelled as the workspace spells them.

Output shape

A one-line "wrote/updated CLAUDE.md" confirmation, a 4-6 line summary of what was filled from real data, then the short list of guessed/blank items to confirm. If you didn't write (because one existed), show the proposed content and the ask instead.