Startup Audit

Startup Audit

Read an already-built product and give a fast, code-grounded Continue / Pivot / Kill triage call. Ingest the codebase (and optionally a live URL), infer the business model into a Lean Canvas, diff coded reality against the claimed story, and ship a self-contained interactive HTML dossier that opens with the verdict.

This skill is a fast triage, not a deep adversarial verdict. Its Stage 2 is a single-pass per-lens findings sweep — no rebuttal round — so the verdict is an honest first read, not a grilled one. For a consequential call (especially Kill or Pivot), the dossier points to startup-grill for the adversarial confirmation. Every claim is tiered by confidence and pinned to a provenance pointer; overclaiming is the cardinal sin. The verdict is opinion, not advice — the disclaimer ships with every run.

A pure-diligence dossier (evidence only, no verdict) is available on request — see Phase 0 mode detection.

STOP — When NOT to use this skill

Hand off — do not run the audit — if any of these apply:

• The product doesn't exist yet (idea / pre-build) → use startup-launch-kit or validation-canvas. There is no codebase to read; this skill would have nothing to infer from.
• The product exists, but the user wants the full 5-artifact kit (brand → canvas → RAT → deck → grill), not a fast verdict → use startup-launch-kit. This skill produces only the Continue/Pivot/Kill triage; the kit is the full-pipeline path (and it reuses this skill's mode=diligence as its code-reader — see Cross-Skill Integration). The split is by output: want the verdict → here; want the kit → there.
• The user wants a DEEP adversarial verdict on belief artifacts (a Lean Canvas or pitch deck), or wants to confirm a consequential Kill/Pivot this skill flagged → use startup-grill. This skill gives the fast code-grounded triage; grill gives the heavier 3-round adversarial verdict on text artifacts. They are siblings split by input (built code/URL vs belief artifacts) and rigor (triage vs adversarial).
• The user wants deep UI/UX review of an AI feature → use ai-ux-review (and ai-eval-review for the eval layer). This skill delegates the UI/UX layer to those skills rather than rebuilding it.
• The user wants to audit a SKILL.md's rule adherence → use skill-evaluator. Different "audit" — that one reads skill text, this one reads a product.

Bright-line rule: if there is no built artifact to read (no repo, no live URL), this is not the skill — pre-build work routes to startup-launch-kit. If the user wants the deep adversarial verdict on a canvas/deck, that's startup-grill.

Redirect discipline. When the STOP gate fires and you route to startup-grill or team-composer, describe the kind of lens needed — don't coin role tags. The canonical persona catalog is skills/team-composer/references/role-personas.md; defer naming specific roles to those skills.

Skill Boundaries

The split with startup-grill is input + rigor, not evidence-vs-verdict: this skill reads a built artifact and gives a fast triage verdict; grill reads belief artifacts and gives a deep adversarial verdict. When this skill returns a Kill or Pivot, it recommends grill for the heavier confirmation — that's a handoff, not a boundary the user has to manage.

What this skill produces

Every run produces files inside the resolved audit root (see Phase 0 for path resolution; default docs/audit/ for solo runs, docs/startup-kit/audit/ when invoked under the kit root):

1. startup-audit.html — the single self-contained interactive dossier. In default (verdict) mode it opens with the Continue / Pivot / Kill verdict
   • Red/Amber/Green band + the disclaimer, then the evidence. In diligence-only mode the verdict section is omitted (evidence only). Format + template: references/dossier-html-template.md.
2. startup-audit.md — canonical, editable Markdown mirror of the dossier.
3. inferred-canvas.md — the inferred Lean Canvas using validation-canvas's exact headings, every field carrying a provenance pointer + confidence tier.

The handoff bridge (read carefully — this is where the chain connects). Downstream skills do not read inferred-canvas.md: startup-grill and pitch-deck grep <canvas-root>/validation-canvas.md, and riskiest-assumption-test reads its own assumption-test-plan.md. So inferred-canvas.md is this skill's evidence artifact, not an auto-consumed pipeline input. The bridge is the offer-to-seed step (Phase 2): if no founder validation-canvas.md exists, this skill offers to seed one — and that seeded file is what grill consumes when the user takes a Kill/Pivot for confirmation. For riskiest-assumption-test, the handoff is a pointer: the dossier recommends running RAT on the unknown blocks. State this honestly — do not imply an automatic feed that isn't wired.

Existing files from prior sessions are appended-to / re-rendered, never silently overwritten. Never overwrite a founder-authored validation-canvas.md — see Phase 2.

Phase 0: Path resolution + mode + dependency pre-flight

Resolve the audit root once, in this precedence order:

1. Explicit output_dir arg → use as-is.
2. STARTUP_KIT_DOCS_ROOT env var set → ${STARTUP_KIT_DOCS_ROOT}/audit/.
3. docs/startup-kit/ exists → docs/startup-kit/audit/ (surface the smart-default notice: "Writing to docs/startup-kit/audit/. Set STARTUP_KIT_DOCS_ROOT=./docs to write standalone instead.").
4. Solo fallback → docs/audit/.

Sibling reads (<canvas-root>/, <rat-root>/) resolve as siblings of the audit root, matching startup-grill's chain.

Mode detection. Default mode emits the verdict. Switch to diligence-only mode (evidence dossier, no verdict / no score / no disclaimer-as-opinion) when the user asks for it — phrases like "diligence only", "evidence only", "no verdict", "just the dossier / no opinion", or an explicit mode=diligence arg. Record which mode ran; the dossier states it.

Dependency pre-flight (fail loud). Before doing any work, check that the required sibling skills are present (see Dependencies). If a required reference is missing, refuse with a clear message — do not silently degrade:

"startup-audit requires team-composer, validation-canvas, and riskiest-assumption-test to be installed (it reads their reference files and writes artifacts they consume). One or more is missing — install the full agent-skills plugin, not just this skill."

Capability gate (degrade with notice). Detect optional runtime capabilities and pick the lane; record which lane ran so the dossier can state it:

• SocratiCode MCP present (mcp__socraticode__codebase_*) → use it for the semantic map. Absent → glob/grep fallback over high-signal files.
• WebFetch / Playwright present AND user supplied a URL AND URL fetch is explicitly opted-in → fetch the live surface. Absent / not opted-in → codebase-only; note the URL was skipped. If a fetch is attempted but fails (timeout, auth wall, error) → fall back to codebase-only, record the failure in the lane note, and continue — a failed URL fetch never blocks the audit.

Phase 1: Signal extraction (Stage 1a)

Goal: sweep the codebase for deterministic business signal. This is a structured extraction, not a judgment call — precision lives here.

Apply references/signal-extraction.md end-to-end. Detect the ecosystem (JS/TS, Python, or generic fallback) and run the matching extractors over the universal source taxonomy:

dependency manifests · data model / schema · routes / pages · auth & tenancy · .env(.example) · money code · README / docs claims (+ roadmap markers) · commit recency / contributor count.

Secret-redaction rule (hard). Extract key names / presence only — never values. A STRIPE_SECRET_KEY entry in .env.example is signal that Stripe is wired; its value (if ever present) must never be read into any artifact. If a real secret is encountered, record only "" and move on.

Output of this phase is a structured signal set (internal), each signal tagged with its source file/path — that tag becomes the provenance pointer in Phase 2.

Phase 2: Inference (Stage 1b) → inferred canvas + build-vs-claim diff

Goal: turn signals into an inferred business model, tiered by confidence, and surface the gap against the claimed story.

Apply references/inference-mapping.md:

1. Infer into the nine Lean Canvas blocks using validation-canvas's exact headings: ### Problem, ### Customer Segments, ### Unique Value Proposition, ### Solution, ### Channels, ### Revenue Streams, ### Cost Structure, ### Key Metrics, ### Unfair Advantage. (These headings are grepped by startup-grill and riskiest-assumption-test — match byte-for-byte.)
2. Derive the confidence tier from provenance — do not judge it:
   • observed — a deterministic signal is directly present (e.g. an invoice.paid webhook handler + a populated plan table).
   • inferred — reasoned from a pattern (e.g. a Stripe dep + a Subscription entity → probably recurring SaaS).
   • unknown — no signal. The block renders empty.
   • Hard gate: provenance == null → the field cannot render as a claim. No provenance pointer, no claim. Unknowns stay unknown; they are not filled by reasoning about the market.
3. Build-vs-claim diff (the headline evidence). Contrast the inferred (coded) model against the claimed story (README / marketing copy / URL). Run it bidirectionally: claimed-but-not-built AND built-but-not-claimed. Parse README roadmap markers ("Future", "Coming soon", "Roadmap") — a feature the tagline sells but the roadmap marks "Future" is a diff finding.

Write inferred-canvas.md to the audit root. The blocks code evidences poorly — typically Problem, Unique Value Proposition, Unfair Advantage — render as unknown and become the riskiest-assumption-test handoff (these are exactly the beliefs a built artifact can't prove).

Thin / greenfield repos are a valid result, not a failure. A sparse repo (prototype, few deps, no schema) legitimately yields a mostly-unknown canvas. Still run the audit — report the unknowns honestly, downgrade the verdict's confidence accordingly (see Phase 3), and route the unknowns to riskiest-assumption-test. Do NOT pad the canvas with market reasoning to make it look complete; that trips the provenance gate. Yield scales with repo maturity.

Never overwrite a founder-authored validation-canvas.md. Write to the separate inferred-canvas.md. If no validation-canvas.md exists at the canvas root, offer to seed one from the inferred canvas (the founder can then correct the machine's inference) — but only on explicit confirmation.

Orchestrated carve-out (no double-prompt): when invoked with mode=diligence by startup-launch-kit (existing-project mode), the founder's Phase 0.3 opt-in ("read the code and pre-fill the canvas") is the seed confirmation — write the seed directly, do not re-prompt. Standalone invocation still requires the explicit confirmation above.

Seed-header contract (load-bearing — do not drop). The seeded validation-canvas.md MUST retain the machine-inferred provenance header from inferred-canvas.md verbatim (the full blockquote in references/inference-mapping.md), plus each block's _(TIER — provenance)_ tag.

The header opens with the detection marker — the HTML comment <!-- SEED:machine-inferred --> — which validation-canvas keys on to run its tiered confirm (observed glance / inferred verify / unknown interview) instead of treating the seed as founder truth. This marker is the single source of truth (defined in references/inference-mapping.md): single-line, invisible when rendered, and absent from any founder-authored canvas. validation-canvas and startup-launch-kit/references/state-detection.md detect the seed by matching this exact comment. If you ever change the marker, update both readers in lockstep — a seed whose marker no longer matches reads as a founder-authored canvas downstream and silently launders machine guesses into belief (the exact failure the provenance gate prevents). This matters most when startup-launch-kit drives the seed in existing-project mode.

Phase 3: Audit panel + verdict synthesis (Stage 2)

Goal: apply domain-aware lenses to the inferred canvas + diff, then synthesize a fast triage verdict. This is a single-pass per-lens findings pass, NOT a multi-round debate. The adversarial opening/rebuttal/synthesis panel is startup-grill's job — do not rebuild it here. The verdict here is an honest first read derived from that single pass; consequential calls route to grill for confirmation.

Apply references/audit-panel-resolution.md, then references/verdict-and-scoring.md:

1. Pre-fill team-composer Phase 1 signals from the inference (not from a brief): tenancy model → audience; compliance configs / regulated entities → is_regulated; ML/LLM deps → is_data_intensive + AI-feature flag; i18n / locale dirs → is_international; voting / nudge / gamification code → involves_behavior_design; commit history / maturity → stage.
2. Resolve the lens panel using the selection algorithm in audit-panel-resolution.md, which reads personas from skills/team-composer/references/role-personas.md (capability-gated; generic lens fallback if absent). Do not invoke team-composer as a sub-skill — read its catalog, mirror startup-grill's read-then-select pattern.
3. Each composed lens emits exactly one findings block against the inferred canvas + diff — focused, evidence-anchored, no debate. Every finding cites the canvas field / diff row / signal it rests on, and is tagged with a finding-id (F1, F2, …) and a severity×fixability per verdict-and-scoring.md.
4. AI features detected → add the AI-safety lens AND route to the conditional skills ai-ux-review (human-AI design + integrity surface) and ai-eval-review (eval rigor). If they run, embed/link their docs/ai-ux/*.html outputs into the dossier rather than re-assessing AI quality here.
5. Options the evidence suggests. Generate grounded options (including pivot directions). Hard gate: every option must cite the finding that motivates it — no citation, the option is suppressed. When the verdict is Pivot, these options ARE the code-grounded pivot directions.
6. Verdict synthesis (default mode only; skipped in diligence-only mode). Per references/verdict-and-scoring.md, derive from the tagged findings:
   • a headline verdict — Continue / Continue-with-conditions / Pivot / Kill (the headline carries the recommendation; never let the band stand in for it);
   • a Red / Amber / Green heat band (coarse, layered under the headline);
   • an evidence-confidence label — a Kill/Red resting on majority-unknown signals downgrades its own confidence and says so;
   • a headline reason that cites specific finding-ids (no verdict without a finding pointer);
   • for Kill or Pivot, the explicit handoff: "This is a fast code-grounded read — run startup-grill for the deep adversarial confirmation." The verdict is opinion — render the disclaimer block from verdict-and-scoring.md with it, and never use a banned term ("valuation" / "guarantee" / "will succeed/fail").

Phase 4: Render the dossier

Apply references/dossier-html-template.md. Write the canonical startup-audit.md, then render startup-audit.html from it.

Dossier sections, in order:

1. Verdict (default mode only — omitted in diligence-only mode) — Continue/Pivot/Kill headline + R/A/G band + evidence-confidence + headline reason citing findings + (for Kill/Pivot) the grill-confirm handoff + the disclaimer block.
2. Executive summary — what this product is, in evidence terms.
3. Build-vs-claim diff — two columns (coded reality vs claimed story); the headline finding sits at the top.
4. Inferred Lean Canvas — nine blocks, each color-coded by confidence tier, each field with an expandable provenance pointer.
5. Audit findings — per-lens findings blocks (with finding-ids + severity).
6. Options the evidence suggests — each option linked to its source finding.
7. Handoff & next steps — startup-grill to confirm a Kill/Pivot; riskiest-assumption-test for the unknown blocks; which optional skills ran.

The HTML is self-contained (inline CSS + vanilla JS, zero network deps, base64 any images, @media print clean) with interactivity: expandable provenance per field, confidence-tier filter, collapsible blocks. Brand tokens injected from DESIGN.md if present.

Run the verifier checklist before presenting:

• Every rendered canvas field has a provenance pointer (or renders as unknown)
• inferred-canvas.md headings match validation-canvas's spec byte-for-byte
• If a validation-canvas.md seed was written: it retains the <!-- SEED:machine-inferred --> marker verbatim on its second line (else downstream reads it as founder-authored and launders the inference)
• Build-vs-claim diff ran bidirectionally
• Every "Options" entry cites a finding (uncited options suppressed)
• Default mode: verdict is one of Continue / Continue-with-conditions / Pivot / Kill, cites ≥1 finding-id, carries an R/A/G band + evidence-confidence, and ships the disclaimer; Kill/Pivot includes the grill-confirm handoff
• Default mode: a Kill/Red on majority-unknown evidence self-flags low confidence
• No banned term ("valuation" / "guarantee" / "will succeed/fail") appears
• Diligence-only mode: no verdict / band / score rendered
• No secret values written anywhere in the artifacts
• HTML opens offline (no network requests) and prints clean
• The lane + mode that ran is stated

If any box fails, fix before presenting. Then present the files.

Quality bars (skill-wide)

The skill must refuse to ship if any of these are true:

• A canvas field claims something with no provenance pointer — provenance or unknown, never a bare assertion.
• An "Options" recommendation cites no finding — that's blind advice; suppress it.
• (Default mode) a verdict appears with no finding-id citation — a verdict without support is rhetoric; cite the findings that drive it.
• (Default mode) the R/A/G band is presented AS the recommendation instead of the Continue/Pivot/Kill headline — the headline carries the call; the band is coarse heat only.
• (Default mode) a Kill/Red on majority-unknown evidence that does NOT self-flag low confidence — the verdict must inherit the evidence quality.
• (Default mode) the disclaimer block is missing, or a banned term appears ("valuation" / "guarantee" / "will succeed/fail") — the verdict is opinion, and must say so.
• A secret value (not just a key name) appears in any artifact.
• The build-vs-claim diff ran only one direction — under-marketed built capability is as much a finding as over-claimed unbuilt capability.

Dependencies

Declared explicitly (unlike most repo skills, which discover deps implicitly) so a per-skill install fails loud rather than mysteriously.

Required skills (fail loud if absent — see Phase 0 pre-flight):

• team-composer — Phase 3 reads references/role-personas.md for lens personas.
• validation-canvas — Phase 2 infers into its exact Lean Canvas headings; the inferred canvas is consumable only because it mirrors this skill's spec.
• riskiest-assumption-test — the unknown-block handoff target.

Conditional skills (invoked on signal; skip + note if absent):

• startup-grill — recommended for the deep adversarial confirmation of a Kill/ Pivot. The verdict vocabulary in references/verdict-and-scoring.md is aligned to grill's kill-report.md labels but is self-contained (no runtime read), so grill's absence never blocks this skill's verdict.
• ai-ux-review + ai-eval-review — invoked only when AI features are detected.

Runtime capabilities (capability-gate WITH fallback; absence is normal):

• SocratiCode MCP → glob/grep fallback.
• WebFetch / Playwright → codebase-only (drop URL input).

Install note: install the full agent-skills plugin, not just this skill. A per-skill npx install copies only this folder and breaks the cross-skill reference reads above.

Cross-Skill Integration

Principle: this skill owns the fast, code-grounded triage verdict — read a built product (codebase/URL) and call it Continue / Pivot / Kill, with every claim provenance-tiered and the verdict framed as opinion. It is the light front-of-line read; startup-grill is the heavy adversarial confirmation.

Graceful degradation: with SocratiCode absent it falls back to glob/grep; with no URL it runs codebase-only; with a thin / greenfield repo it honestly reports mostly unknown, downgrades the verdict's confidence, and routes to validation-canvas + riskiest-assumption-test (yield scales with repo maturity — a thin repo can't support a confident verdict, and the band says so rather than faking certainty).