Startup launch kit

Startup Launch Kit

Opt-in orchestrator that sequences the five-step startup pipeline with state tracking via kit-manifest.json. The orchestrator is convenience for founders who want end-to-end coordination without manual step-chaining; it is not a replacement for any individual skill, and it never bypasses the gates each skill enforces.

The orchestrator's job is sequencing + state, not content. Each pipeline skill runs as itself — same prompts, same outputs, same gates the founder would see if they invoked the skill directly. The orchestrator just remembers where the founder is, surfaces the next step, and records overrides for audit.

It runs in one of two source modes (detected in Phase 0.3): greenfield (no built product — the step-by-step pipeline, as always) or existing-project (a codebase is present — the orchestrator reads it via startup-audit and pre-fills what code reveals, asking the founder only for what it can't infer). Existing-project mode is opt-in: offered when a codebase is detected, never forced.

---

STOP gate — When NOT to use this skill

Hand off to a different skill — do not run the orchestrator — if any of these apply:

• The user wants only one step, named explicitly (e.g., "just build my validation canvas", "rework my pitch deck slide 5", "grill my idea") → route to the named skill directly. The orchestrator's value is end-to-end sequencing; running it for one step is overhead with no benefit.
• The user is mid-step inside one skill, asking for changes within that skill's scope (e.g., "update the UVP block", "re-rank the top-3 hypotheses") → route to the same skill in update mode. The orchestrator doesn't add value to within-skill iterations.
• The user wants to discuss the pipeline conceptually ("which validation method should I use?", "explain how the gates work") → answer directly or route to team-composer with @startup_strategist. The orchestrator executes; it doesn't teach.
• The user is grilling someone else's startup (advisor / investor mode) → route to startup-grill directly. The orchestrator assumes the founder is building their own pipeline.

Bright-line rule: if the founder names a single pipeline skill they want to invoke, route to that skill. The orchestrator only runs when the request is genuinely end-to-end ("the whole pipeline", "everything from brand to grill", "set up my launch kit"), or when the founder explicitly invokes /startup-launch-kit.

---

What this skill produces

The orchestrator does not produce its own content artifacts — it produces sequencing + state. Specifically:

1. docs/startup-kit/kit-manifest.json — a thin JSON state journal: manifest version, current step, completed steps with mtimes, gate-override flags with reason+timestamp, founder-intake-answers cache. Atomic writes (write .tmp, then rename). The orchestrator creates the docs/startup-kit/ folder if absent. See references/manifest-schema.md for the schema and three worked examples.
2. Sequenced invocations of the five pipeline skills via the Skill tool. Each skill runs as itself with the founder seeing its normal prompts; the orchestrator merely calls them in order, gates the transitions, passes the per-skill output_dir (docs/startup-kit/<skill>/), and reads the manifest entries each skill writes after completing.
3. A final-summary report when all five steps complete: artifact list, gate-override log, iteration count, time-elapsed.

The orchestrator does NOT:

• Re-implement any skill's content (no canvas blocks, no test methods, no slide construction, no kill-report logic).
• Bypass any gate enforced by any pipeline skill.
• Batch multiple steps without surfacing each step's prompts.
• Auto-route loop-back actions (it surfaces recommendations; the founder decides).

---

Phase 0: STOP gate + path resolution + manifest discovery

Apply the STOP-gate rules above first. If the request passes, then:

Step 0.1 — Resolve the kit root

Resolve the kit root once at invocation, in this precedence order (canonical chain):

1. STARTUP_KIT_DOCS_ROOT env var → ${STARTUP_KIT_DOCS_ROOT}/startup-kit/
2. Default → docs/startup-kit/

The orchestrator passes <kit_root>/<skill-folder>/ to each child skill as its output_dir argument when invoking. Per-skill folder names (brand/, canvas/, rat/, pitch/, grill/) are fixed by the conventions doc.

If the kit root directory does not exist yet, create it before manifest discovery (mkdir -p semantics).

Step 0.2 — Manifest discovery

1. Look for kit-manifest.json at <kit_root>/kit-manifest.json first.
2. Backward-compat fallback: if absent, also check the working-directory root (./kit-manifest.json). If found at the legacy path, surface a one-line notice so the founder knows: "Found legacy kit-manifest.json at repo root. To consolidate, mv kit-manifest.json docs/startup-kit/ after this run." Continue using the legacy location for this run; do NOT auto-move.
3. If absent in both locations: this is a fresh start. Create the manifest at <kit_root>/kit-manifest.json at the end of Phase 1 once intake answers are captured. Proceed to Phase 1.
4. If present: read it. Reconcile against the filesystem per references/state-detection.md:
   • Files exist that the manifest doesn't list → assume manual run; absorb into the manifest with status: "completed" and the file's mtime.
   • Manifest references files that don't exist → mark step as failed; recommend re-running that step.
   • Manifest's intake-answers cache exists → surface them and ask the founder to confirm or update before proceeding to Phase 2 (skip re-asking from scratch).

Step 0.3 — Source-mode detection (greenfield vs existing-project)

The orchestrator runs in one of two source modes. Detect, then offer — never force.

Resume precedence — check this FIRST, before evaluating the offer trigger below. If a machine-inferred seed already exists at the canvas root (the seed marker <!-- SEED:machine-inferred --> per references/state-detection.md Rule 7), do NOT re-offer and do NOT re-invoke startup-audit — a prior run already read the code. Skip Phase 0.6 and let Phase 2 route the validation-canvas step to confirm-inferred-seed mode. The offer trigger below fires only when no canvas of any kind exists yet.

Greenfield (default). No built product to read. Run the pipeline step-by-step as today. This is the path when the working directory has no codebase, or the founder declines the existing-project offer.

Existing-project (opt-in). The founder already built the product; the codebase encodes answers they shouldn't have to re-type. Trigger the offer when BOTH hold:

1. A codebase is present in the working directory — detect using the same signal categories startup-audit keys on (dependency manifest, data-model / schema, routes/pages, .env(.example)). Do NOT re-specify the detection taxonomy here — see startup-audit/references/signal-extraction.md. A surface check (manifest file exists) is enough to offer; full extraction is startup-audit's job in Phase 0.6.
2. No founder-authored validation-canvas.md exists at the canvas root (<kit_root>/canvas/validation-canvas.md or a legacy location). A canvas carrying the machine-inferred seed marker (<!-- SEED:machine-inferred -->) is already handled by the resume-precedence guard above — it returns early, so a prior seed never reaches this condition.

When both hold (and no seed exists yet, per the resume-precedence check above), surface a one-line opt-in offer:

"I see a codebase here but no validation canvas. Want me to read the code first and pre-fill what it reveals (you confirm or correct), then run the rest of the kit? Or start the full step-by-step pipeline from scratch?"

• Founder accepts → set source_mode: "existing-project" in the manifest; proceed to Phase 0.6.
• Founder declines, or no codebase → set source_mode: "greenfield"; skip Phase 0.6 and proceed to Phase 1 as today.

The offer is opt-in by design: greenfield founders' step-by-step learning path must never be hijacked by the mere presence of a repo.

---

Phase 0.6: Code read — reuse startup-audit (existing-project mode only)

Skipped entirely in greenfield mode. When source_mode is existing-project, the orchestrator reads the codebase by invoking startup-audit, never by reimplementing extraction.

1. Invoke Skill(startup-audit) with mode=diligence and output_dir=<kit_root>/audit/. Diligence mode produces the inferred Lean Canvas (tiered observed/inferred/unknown, provenance-pinned) and the build-vs-claim diff, with no Continue/Pivot/Kill verdict — startup-grill remains the kit's single terminal verdict authority.
2. Audit seeds the canvas (no re-prompt). startup-audit writes inferred-canvas.md to the audit root and seeds <kit_root>/canvas/validation-canvas.md from it — the founder's Step 0.3 opt-in is the seed confirmation, so startup-audit's orchestrated carve-out writes it without a second prompt (it still never overwrites a founder-authored canvas). The seeded file carries the machine-inferred seed marker (<!-- SEED:machine-inferred -->) and per-block _(TIER — provenance)_ tags — that marker is what validation-canvas keys on in Phase 2 to run its tiered confirm (not a blank interview). Verify the seeded file exists before continuing; if it is absent, treat that as the no-seed case and apply step 3.
3. No-seed fallback. If startup-audit produces no seed for any reason — unavailable in this runtime, refused on its own dependency pre-flight, errored, or wrote no validation-canvas.md — fall back to greenfield with a one-line notice (e.g. "Couldn't read the codebase automatically; running the step-by-step pipeline instead.") and set source_mode: "greenfield". Do NOT reimplement signal extraction inline.

Then proceed to Phase 1. When a seed was written, it is consumed in Phase 2 as the orchestrator reaches the validation-canvas step (which detects the sentinel and runs confirm-inferred-seed mode).

---

Phase 1: Intake (once, written to manifest)

Run the same three calibration questions that validation-canvas Phase 0.1–0.3 asks:

1. "Have you founded or co-founded a startup that reached paying customers before?"
2. "Is this idea in a domain you've worked in professionally?"
3. "Do you have direct experience with this customer segment?"

Map the answers to a mode (Guided / Focused / Compressed-with-Challenge per the same matrix as validation-canvas). Confirm the inferred mode in one line. Write the answers to the manifest's intake_answers cache.

Note: this duplicates the question set in validation-canvas Phase 0 deliberately. The orchestrator asks at pipeline start so downstream skills can read the cache and offer to confirm/update rather than re-asking blind. The cache is a hint, not a bypass — validation-canvas Phase 0 still runs in full, just with cached defaults.

If the manifest already had intake answers (resume path), present them and offer confirm/update — do NOT silently re-use without confirmation. The founder's experience may have changed; the manifest is a hint, not truth.

---

Phase 2: Sequence execution

For each step in pipeline order — brand-workshop, validation-canvas, riskiest-assumption-test, pitch-deck, startup-grill — do the following:

Step 2.1 — Manifest check

Read the current manifest entry for this step:

• status: "pending" → fresh run; proceed to Step 2.2.
• status: "completed" → ask the founder: "Already completed on [date]. Update mode (re-run with prior artifacts), fresh run (rebuild from scratch), or skip to next step?" Then proceed accordingly.
• status: "blocked" → the prior orchestrator run hit a gate STOP. Surface the blocker (blocked_by) and route the founder to resolve it (typically: re-run the upstream skill in update mode).
• status: "failed" → an artifact disappeared between runs. Recommend re-running this step.

Step 2.2 — Invoke the skill

Use the Skill tool to invoke the pipeline skill by name (e.g., Skill(brand-workshop)). The orchestrator does NOT inline the skill's content. The skill runs as itself, surfaces its own prompts to the founder (intake questions, draft canvases, slide interviews, defense rounds), and when it completes it writes its own manifest entry per the Phase A pattern each pipeline skill carries.

The founder sees the same UX they would see invoking the skill directly. The orchestrator's role is purely sequencing — call the skill, wait for it to finish, read the updated manifest, decide what comes next.

Step 2.3 — Gate enforcement

After the called skill completes, read its manifest entry. If the entry records a downstream gate STOP (e.g., RAT Results empty for top-3 hypotheses → blocks pitch-deck), surface the STOP message verbatim and pause the orchestrator. The founder can resolve the block manually (re-run the upstream skill in update mode, fill in the missing data, re-invoke the orchestrator) OR explicitly accept an override (recorded in manifest per references/gate-override-protocol.md).

The orchestrator never silently bypasses a gate. The whole point of the orchestrator preserving the pipeline philosophy is that gates remain visible and overrides remain audited.

Step 2.4 — Move to the next step

Once gates are satisfied (or overridden with reason), advance to the next pipeline step. Loop back to Step 2.1 with the new step's manifest entry.

---

Phase 3: Loop-back surfacing

After RAT Results land OR after startup-grill ships its kill report, scan for invalidated hypotheses or canvas-block weaknesses:

• From RAT: any top-3 hypothesis with result: "invalidated" in the assumption-test plan (<kit-root>/rat/assumption-test-plan.md or legacy rat/assumption-test-plan.md).
• From grill: any "Lethal & Fixable" or "Material & Fixable" weakness whose Suggested attack line names a canvas block.

If found, surface a one-line recommendation to the founder, not an auto-route:

"Hypothesis #2 was invalidated (per-seat pricing failed). That points to the Revenue Streams block of your validation canvas. Re-invoke validation-canvas in update mode? (Y/n)"

The founder decides. Loop-back stays founder-driven — the orchestrator recommends, never executes a loop-back without explicit confirmation. This preserves the "sequential teaches iteration" philosophy: the founder sees that invalidation happens, and they choose to act on it.

---

Phase 4: Final summary

When all five steps record status: "completed" AND no unresolved loop-back recommendations remain, surface a final summary:

## Startup Launch Kit — Complete

**Pipeline run summary**

| Step | Status | Iterations | Last updated |
|---|---|---|---|
| brand-workshop          | ✅ completed | 1 | YYYY-MM-DD |
| validation-canvas       | ✅ completed | 2 | YYYY-MM-DD |
| riskiest-assumption-test | ✅ completed | 1 | YYYY-MM-DD |
| pitch-deck              | ✅ completed | 1 | YYYY-MM-DD |
| startup-grill           | ✅ completed | 1 | YYYY-MM-DD |

**Artifacts produced (under `docs/startup-kit/`):**
- audit/ (inferred-canvas.md + diligence dossier — existing-project mode only)
- brand/ (logo, brief, design system, descriptions, banners)
- canvas/validation-canvas.md + .html
- rat/assumption-test-plan.md + test-matrix.html
- pitch/deck.html + speaker-notes.md + deck-checklist.md
- grill/kill-report.md

**Gate overrides recorded:** N (see docs/startup-kit/kit-manifest.json `gate_overrides[]`)

**Time elapsed:** [start] → [end]

**Terminal verdict:** see docs/startup-kit/grill/kill-report.md

Point the founder to docs/startup-kit/grill/kill-report.md as the authoritative read on where the startup stands.

---

Hard rules

1. Independent invocability is sacred. Every pipeline skill must work identically when invoked directly without the orchestrator. The orchestrator's edits to those skills (Phase A of v2.1.0) are additive manifest-awareness only; no skill depends on the orchestrator existing.

2. Gates are never bypassed. The orchestrator can advance past a STOP only via an explicit, recorded override in gate_overrides[] with a founder-acknowledgment flag. Silent bypass is forbidden.

3. No batching. Each pipeline skill surfaces its own prompts. The orchestrator does not collect all intake up front and then auto-feed each skill — that would defeat the value of each skill's discipline.

4. Loop-back is founder-driven. The orchestrator recommends; never executes a loop-back without explicit confirmation.

5. Manifest is a hint, never truth. When manifest disagrees with filesystem, the filesystem wins (real artifacts override stale manifest entries). When intake-cache disagrees with current founder state, ask.

6. No content duplication. The orchestrator never grep-parses heading anchors that the called skills own. It reads only the manifest's structured fields. If a downstream check needs content from an artifact (e.g., pitch-deck checking RAT Results), the downstream skill does that read — not the orchestrator.

7. Single working directory, single kit root. The orchestrator assumes one working directory for the whole pipeline AND a single kit root (docs/startup-kit/ by default, or ${STARTUP_KIT_DOCS_ROOT}/startup-kit/ if the env var is set). The orchestrator passes the resolved per-skill output_dir to each child skill at invocation. Multi-directory composition (multiple startups in one repo) is out of scope; founders running multiple startups should use separate working directories.

8. Inference is reused, never reimplemented. In existing-project mode the orchestrator reads the codebase by invoking startup-audit (mode=diligence) — it never greps a package.json, parses a schema, or duplicates any signal extraction. One inference engine (startup-audit), one provenance contract. If startup-audit is unavailable, fall back to greenfield; do not inline a substitute reader. (Mirrors rule 6's no-content-duplication discipline.)

---

Cross-Skill Integration

Skill	When to Use
brand-workshop (our own)	Step 1 of the pipeline; orchestrator calls it. Independently invocable for brand-only work.
validation-canvas (our own)	Step 2; orchestrator calls it. Independently invocable for canvas-only work or update mode.
riskiest-assumption-test (our own)	Step 3; orchestrator calls it. Independently invocable for test design or results-update.
pitch-deck (our own)	Step 4; orchestrator calls it. Independently invocable for deck construction or single-slide rework.
startup-grill (our own)	Step 5; orchestrator calls it. Independently invocable for adversarial review at any stage.
startup-audit (our own)	Existing-project mode only (Phase 0.6). The orchestrator invokes it with mode=diligence to read the codebase and seed validation-canvas.md (inferred Lean Canvas, tiered + provenance, no verdict). Reused as the code-reader — never reimplemented. Independently invocable for a standalone Continue/Pivot/Kill triage of a built product.
team-composer (our own)	Use instead of the orchestrator for pipeline-strategy discussions, single-block deep dives, or naming/positioning workshops that don't fit the pipeline shape.
theme-factory (Anthropic)	Independent of the orchestrator; can be applied to any artifact for visual styling.
pdf (Anthropic)	Independent; for assembling the orchestrator's output artifacts into a board packet.

Principle: the orchestrator owns sequencing, state, and gate-honoring. Each pipeline skill owns its content and gate enforcement. No skill depends on the orchestrator existing; the orchestrator depends on every pipeline skill being manifest-aware (added in v2.1.0 Phase A).

Graceful degradation: if the runtime does not support invoking the Skill tool from within a skill, fall back to instructional prose at each step boundary: "Now run validation-canvas — I'll wait for the artifact to appear, then continue." The orchestrator can still manage the manifest and gate-checking even if direct skill invocation isn't available.