business-teardown

Tear down one named small business — its online presence, what reviews actually say, who it really competes with, how it prices, why customers can't find it — and come back with prioritized, evidence-

SKILL.md

GitHub

Small business teardown

In the owner's words: "Business is down and I don't know why." · "How do we look next to the place across the street?" · "I'm thinking of buying this salon — what am I walking into?"

Connect the server (once)

The tools live on the free Brick & Mortar Small Business Intelligence MCP server: https://brickandmortar.dev/mcp — Streamable HTTP, no account, no key, no OAuth, read-only.

  • Claude Code: claude mcp add --transport http brickandmortar https://brickandmortar.dev/mcp (add --scope user to have it in every project). Check with claude mcp list.
  • Claude.ai / Claude Desktop: Settings → Connectors → Add custom connector → paste the URL.
  • ChatGPT: search the app directory for "Brick & Mortar Public Records".
  • Any other MCP client (Cursor, VS Code, Windsurf…): add a remote HTTP server with that URL.
  • Raw HTTP: POST JSON-RPC with Accept: application/json, text/event-stream; the reply is an SSE data: frame. The edge refuses a bare script User-Agent, so send a browser-style one.

Install docs: https://brickandmortar.dev/connect/

Call

business_teardown — required business_name (as on its own sign or site, not a guess) and city_metro ("Saint Paul, MN"); optional category ("nail salon"). If you don't pass a category, the procedure confirms it from the business's own menu or service list, because the category Google auto-assigns is often wrong.

What comes back, and the one way to get it wrong

This tool fetches no data. It returns a research framework with these keys: framework (the method, named), research_procedure (numbered steps), output_schema (the shape of the finished deliverable), quality_rubric (good_looks_like, common_failure_modes) and caveats. You execute the procedure with your own web search, then write the deliverable in the output_schema shape and check it against the rubric. Handing the framework back to the user as though it were the finding is the failure this skill exists to prevent.

The method is a four-quadrant teardown: Presence (can they be found), Perception (what do people say), Position (how do they compare), Performance signals (public proxies for trajectory). Each quadrant is scored on its own before any overall verdict, so one bad quadrant doesn't drag down a good one and a strong one doesn't hide a weak one. Recommendations are ranked by evidence strength × likely impact ÷ effort, and every one cites the evidence that motivated it. Before recommending anything that costs the owner money (raise prices, hire, extend hours), the procedure checks local pay and category price inflation in federal series.

Go deeper where the teardown opened a gap

What the teardown surfaced Next tool Skill
The rating hides something; reviews feel off review_intelligence review-intelligence
"Who else is really doing this here?" competitor_landscape competitor-landscape
"Are they priced right?" pricing_benchmark pricing-benchmark
"Why can't anyone find them?" local_visibility_audit local-visibility-audit
"Should I buy it?" broker_diligence_prep broker-diligence-prep
Address is in Minneapolis–St. Paul twin_cities_lookup / twin_cities_records twin-cities-property-records

Rules

  • A rating is not review signal. A high average can mask a bad last ninety days.
  • No public signal shows revenue, margin or financial health. Review velocity and Popular Times are directional proxies; say so rather than implying financial diligence.
  • Verify a chain's reviews are about this location before quoting them.
  • Outside the Twin Cities, every property figure came from research, not a record — say so.

Finish

  • Check the draft against quality_rubric.common_failure_modes before answering.
  • State the caveats that applied. What public signal cannot show is part of the answer.
  • Date-stamp anything that moves (ratings, review counts, prices, multiples).
  • For a client-ready document, pass your finished write-up(s) to compose_report (business_name, audience: owner | broker | buyer | investor | general, completed_analyses: [{tool, summary}]). It returns the section order (verdict first), a contradiction ledger across analyses, and citation rules; you write the report. Pass your finished analysis as summary, never the raw framework.