Business Edge

Business Edge

The analyst-and-mentor layer. It diagnoses a business, decides what matters, and hands specific work to specialists. It does not do the specialist work itself.

Boundary. This skill produces analysis, strategy and a dispatched action list. The actual changes are made by the specialist skills in §6 and executed by bob. If you find yourself writing schema markup or CSS here, you have left your lane.

1. The engagement

P0 SCOPE ─► P1 SCAN ─► P2 DIAGNOSE ─► P3 POSITION ─► P4 PLAN ─► P5 DISPATCH ─► P6 MEASURE ─► P7 LEDGER

Each phase has a required output. Do not skip forward — P3 without P2 produces generic advice, which is the failure this skill exists to prevent.

Roles are data, not skills: roster/*.yaml. Load the relevant role's card when working its phase. Seven roles, zero additional SKILL.md files.

2. P0 — SCOPE

Establish what the business is and how it makes money. Fifteen minutes here prevents a whole engagement of wrong-denominated advice.

2.1 The profit model — ask before anything else

Record every figure as measured / operator-stated / estimated. State assumptions inline and flag the most load-bearing one.

Flagging an assumption does not license concluding from it (S074). Before stating any directional recommendation, re-run it at the plausible range of every assumed input. If the direction flips anywhere in that range, you do not have a recommendation — you have an open question, and it must be reported as one.

This is not hypothetical. A real engagement assumed a 15% gross margin, correctly flagged it as the single most load-bearing assumption, and still concluded "paid advertising is arithmetically foreclosed" — a conclusion that reverses at 25%. The assumption was labelled and the conclusion was stated anyway, which is how a flag becomes decoration.

UNGROUNDED — direction depends on an assumed input
  claim:      paid acquisition is foreclosed
  hinges on:  gross margin (ASSUMED 15%, never measured)
  flips at:   ~20%
  to resolve: one COGS figure from the operator

Report it in that shape. UNGROUNDED is a first-class outcome, never a weaker PASS: it names the one number that would settle the question, which is far more useful to an operator than a confident answer derived from a guess. Only an input recorded as measured can carry a directional claim on its own.

2.2 Also capture

Goal and time horizon · constraints (budget, operator hours/week, technical capability) · what has already been tried · the competitor set · what the operator believes is true (to be tested, not assumed).

3. P1 — SCAN (all dimensions, no false negatives)

Spawn the scout agent for the off-site half — it owns surface intelligence and enforces the probe rules. Browser-first; APIs are a fallback, never the sole source. Record probe status on every probe via scripts/probe_ledger.py, and run probe_ledger.py coverage before writing any finding. Full method: references/probe-protocol.md.

scout reports what exists and what could not be seen. It does not diagnose or recommend — it hands the surface map back here, where P2 owns diagnosis.

Dimensions — sweep all, mark NOT_PROBED explicitly where skipped:

Also scan what AI says about the business. Ask the major assistants the buying questions a real customer would, and record the answer and its cited sources. The AI's brand summary is now a public artifact the operator does not control and often has never read — it frequently contains their defects verbatim.

4. P2 — DIAGNOSE

4.1 Strengths first — they are assets, not boxes ticked

Most audits only find faults. That is a defect: the thing worth compounding is usually already working and invisible on a gap list. Name each strength, why it exists, and whether a competitor could copy it. Founder-level expertise, response speed, and genuine technical specificity are routinely the most valuable and most under-used assets a small operator has.

4.2 Objections — what actually blocks the sale

Mine real objections from calls, reviews, community threads, support tickets, configurator drop-off and abandonment — not from keyword tools, which only see demand that already exists. For each: is it resolved anywhere a buyer or an AI can find, and who custodies that resolution (you, or a third party)?

4.3 Journey gap map

Map every finding to found → selecting → buying → after-sales, then identify the binding constraint — the one stage where fixing anything else changes nothing. Say which it is and why.

5. P3 — POSITION

5.1 Prefer asymmetry over best-practice

An action qualifies as edge only if a larger competitor cannot easily copy it — usually for organisational or economic reasons rather than technical ones. Incumbents can copy any tactic; what they cannot do is justify small, specific, unscalable work.

Test each candidate: could a competitor with 50× the budget do this tomorrow? If yes, it is hygiene — do it if cheap, but never call it strategy.

Structurally asymmetric for a small operator: publishing verifiable operational specifics · correcting a defect publicly and fast · genuine named-expert participation · direct human access at the moment of doubt · learning latency (days, not quarterly cycles) · owning narrow segments an incumbent finds economically trivial.

Commodity theatre — reject by default: generic content calendars · "get more reviews" as strategy · schema markup as a moat · posting-frequency retainers · dashboards with no intervention attached · cosmetic CRO while speed/clarity/payment are broken · paid retargeting labelled as edge.

5.2 Never propose manipulation

Creating an artifact is legitimate. Manufacturing apparent independence is not. Required: identity and commercial relationship disclosed where the answer appears · a real pre-existing question · falsifiable claims backed by records · first-party statements clearly separated from independent corroboration · correction when facts change · standalone usefulness even if no AI ever retrieves it.

Prohibited: fake personas · planted questions · undisclosed compensation · vote rings · fabricated customers · presenting the operator's own claims as community consensus. Beyond the ethics: fake signals are removed and take their citations with them, so the tactic is self-erasing.

5.3 The verification path — usually unmanaged, usually decisive

Design what happens when a prospect leaves to check whether the business is real. Route them to records the operator does not control — marketplace transaction history, independent review platforms, company registration, dated evidence. Third-party custody is what makes a claim credible to both humans and retrieval systems; a policy page on your own domain is not corroboration.

6. P4/P5 — PLAN and DISPATCH

Every action gets: owner skill · effort · expected effect in PROFIT terms · evidence grade · kill criterion. Rank by (profit impact ÷ operator-hours), never by revenue.

Dispatch map — action type → specialist

If no specialist owns an action, say so explicitly and file it as a library gap. Do not silently absorb specialist work into this skill — that is how a 240-line generalist ends up owning five domains it cannot maintain.

7. P6 — MEASURE

7.1 Check statistical feasibility BEFORE designing any experiment

Compute what the order volume can actually detect. At single-digit orders per month, order-level A/B testing is impossible — a two-arm conversion test can need years per arm. Saying so is mandatory; proposing a test that cannot conclude is a serious error.

When order volume is too low, move the measurement unit up the funnel to something with hundreds of observations per month — AI-answer composition, citation presence, impressions, quote requests. Orders remain ground truth for profit; they are simply unusable for significance.

7.2 Measure to profit, and instrument attribution honestly

Track exposure → visit → enquiry → order → profit after channel fees → returns. Where a large share of sales is untagged, the cheapest fix is a post-purchase "how did you first hear about us?" — zero-party data, and the only instrument that sees zero-click discovery.

8. P7 — LEDGER

Record every action, its prediction, its result — including nulls and adverse effects — and its evidence grade. Promote a method only on replication:

conjecture → single-entity signal → replicated entity result → cross-category pattern

Failed methods stay recorded with their failure evidence rather than being deleted, otherwise the library forgets what does not work and proposes it again.

8.1 Keeping the advice itself from rotting

The same decay applies to this skill family. scripts/claims_lint.py finds a fact owned by several skills whose verdicts disagree — the state that precedes a confidently wrong recommendation:

python3 ~/.claude/skills/business-edge/scripts/claims_lint.py drift --show-duplicates
python3 ~/.claude/skills/_meta/gates.py G_CLAIM_FRESHNESS --claim-mode strict   # exit 2 on drift

Duplication alone is reported, never blocking; contradiction blocks in strict mode. Run it after any change to a fast-moving claim. A contested fact needs one owner; the rest point at it.

9. Anti-Patterns