fs-flagship-chain-campaign

name

fs-flagship-chain-campaign

description

The decision-gated executable campaign that proved and hardened the flagship founder chain (market-validation → build-options → prd → tasks). Phases 0-3 executed 2026-07-02 (template rescue, proving runs, prd/tasks rewrite, site truth-up) — the skill remains the record and the playbook for regressions or a future re-run. Use when a maintainer says "run the proving run", "prove the workflow scripts", "harden the founder chain", "is the flagship chain fixed", "rewrite prd/tasks", "genericize agent-browser", "fix the site oversell", or asks what the current phase/status of the flagship campaign is.

Flagship Chain Campaign

You are executing a phased, decision-gated campaign on the repo's hardest live problem. The flagship chain is the four-plugin founder pipeline /market-validation → /build-options → /prd → /tasks (handoffs pass file paths, never inline content). Its execution end is unproven: the two Workflow scripts (JS files run by Claude Code's Workflow({scriptPath, args}) tool, which orchestrates sub-agents) were self-declared "syntax-checked only" until the 2026-07-02 proving runs, and prd/tasks were off-style external imports until the same-day rewrite. A proving run = the first real execution, judged only by measurable observations.

Announce at start: "Running the flagship-chain campaign — current phase status below; every gate is measured (exit codes, counts, shapes), never judged by eye."

Iron rules of this campaign:

1. Promotion of every phase routes through fs-change-control (spec in docs/superpowers/specs/, date-prefixed, operator-approved → checkbox plan in docs/superpowers/plans/ → commits matching the plan verbatim). No phase ships around it.
2. Assistants run no mutating git commands; the operator reviews and commits.
3. Success is MEASURED: test counts, gate-script exit codes, rendered artifacts. Run .claude/skills/fs-flagship-chain-campaign/scripts/gate.sh at every gate.

Campaign status board (as of 2026-07-02)

Phase	What	Status	Gate
0	Template rescue + gitignore fix	DONE — committed 648d304, pushed 2026-07-02	gate.sh phase0 → exit 0 ✓
1	Proving runs of both Workflow scripts	EXECUTED 2026-07-02 (minimum-honest scale, real topic): 1A passed all expected observations (4/4 angles, 24 curated/24 survived, shape+args ✓); 1B ran end-to-end and CAUGHT ledger entry 9 (silent skeptic-verdict drop) — fixed by-construction, stress-test leg re-run	gate.sh phase1 <r.json> <b.json> → exit 0
2	Rewrite prd/tasks into house style + genericize agent-browser	DONE 2026-07-02, uncommitted — all 9 criteria pass (independent acceptance agent)	measurable criteria below ✓
3	Site truth-up (docs/index.html oversell)	DONE 2026-07-02, uncommitted (operator un-gated in-session; 3 maintainer-tooling cards added; site-checks.sh green)	strings removed ✓ + fs-release-and-publish gate

When NOT to use this skill

You actually want to…	Load instead
Understand the repo / find the right skill	fs-orientation
Write/format the spec+plan for a phase's promotion	fs-change-control
Push/publish a completed phase	fs-release-and-publish
Recreate the venv / run or extend the three test suites	fs-toolchain-and-tests
Learn the house style Phase 2 must target	fs-skill-style-guide
Edit the Pages site itself (Phase 3 mechanics)	fs-site-and-positioning
Read the history of how these defects happened	fs-failure-archaeology
Workflow/Agent/AskUserQuestion semantics in general	fs-anthropic-primitives

Phase 0 — Template rescue (DONE 2026-07-02; record + guard)

What happened: plugins/build-options/assets/matrix.template.html was never committed — the plugin .gitignore line *.html swallowed it, so build_matrix.py (which hard-requires it at HERE / "matrix.template.html") shipped broken; 2 pytest failures (FileNotFoundError). The sole surviving copy was rescued from the mutable plugin cache into the repo, and .gitignore gained the negation !assets/matrix.template.html. Full history: fs-failure-archaeology (ledger #1).

State: committed as 648d304 and pushed 2026-07-02 (the handoff library + 3 skill-* plugins + marketplace registration rode the same commit — the manifest references ./plugins/skill-* sources, so they had to ship together on the live-publishing master).

Your job for Phase 0 now: it's a guard, not a task — run the gate after any change touching build-options assets or gitignores.

.claude/skills/fs-flagship-chain-campaign/scripts/gate.sh phase0

Expected: GATE PASS: phase0 (template intact, gitignore negation present, 9+6 pytest, 5 node:test), exit 0. If pytest can't import: the host Python (3.13.5) is PEP-668 externally managed — create the venv per fs-toolchain-and-tests, or pass PY=<venv>/bin/python to the gate script (its default ~/venvs/founder-skills/bin/python is environment-specific).

Phase 1 — Proving runs (the two syntax-checked-only Workflow scripts)

Both scripts read args defensively (JSON string or object) and were only ever executed in an earlier hardcoded form. The unknown being proven is the live Workflow runtime + args threading + return shape, NOT the internal logic (the JS harness already covers the investigator retry logic — re-mocking proves nothing new).

Scale options, ranked (pick with the operator via AskUserQuestion):

1. (Recommended) Minimum-honest run — a real product idea, but minimal knobs: market-validation with only the 4 default angles (pain/competitor/wtp/counter); build-options with 3 lenses and the 8 default criteria. Proves args plumbing + shapes at a fraction of the reference cost. Still real web-research spend.
2. Full-fidelity run — 4 fixed + 4–7 product angles, 6 lenses. Also proves the Phase-2 manual seam at reference cost (~1.5M tokens, ~38 agents, ~50 min on a 4-core Pi, per market-validation SKILL.md).
3. FENCED: mocked "proving" run — stubbing agent() again duplicates the existing node:test suite and proves nothing about the live runtime. Do not count it as the proving run.

1A — market-validation research-workflow.js

Invocation (from the repo checkout; <repo> = the founder-skills root — the installed plugin cache copy is byte-identical at HEAD 2e4c9dd but may desync after future pushes, so prefer the repo path):

Workflow({
  scriptPath: "<repo>/plugins/market-validation/assets/research-workflow.js",
  args: { config: { product, oneLiner, customer, geography, date, scope },
          angles: [ { key, title, focus }, ... ] }   // omit angles → 4 defaults
})

Expected observations, in order (from the script source, verified 2026-07-02):

Gate	Expect
Phases announced	Investigate → Curate → Verify → Synthesize
Agent labels	investigate:<key> per angle (+ investigate-loose:<key> only on strict-schema failure), then curate, then verify:<i> one per curated claim, then synthesize
Args threading	Your product name / scope text appears inside investigator prompts (SCOPE line). Generic "the product" = args did NOT thread
Log after Investigate	Collected N raw claims across M of K dispatched angles (dropped angles named: DROPPED ANGLE "<key>" …)
Log after Curate	Curated N claims; M competitors — N should be 18–24 on a real run
Log after Verify	Verified: X survived, Y dropped
Return shape	{ report (markdown string), survivors[], competitors[], droppedCount }

Then: save the return value verbatim to research-output.json and run gate.sh phase1 research-output.json build-options-output.json (after 1B).

If you see X instead:

• Prompts say "the product" / generic scope → args didn't thread. Check whether the Workflow host passed args as a JSON string vs object (script handles both) — if neither arrived, the args-passing convention itself is broken; record in fs-failure-archaeology and stop.
• curate throws / returns null → the workflow has NO retry outside investigators; a failed curate/verify/synthesize call aborts or empties the run. toVerify becomes [] → 0 survivors. Treat as a failed proving run, not "the market has no evidence".
• All angles DROPPED → allClaims is empty; abort rather than let curate hallucinate. Likely cause: the host's structured-output (schema) support differs from what agent() provides.
• droppedCount very high (> survivors) → verification working as designed but sources weak; the run is still a successful proving of the script — note the distinction.

1B — build-options build-options-workflow.js

Workflow({
  scriptPath: "<repo>/plugins/build-options/assets/build-options-workflow.js",
  args: { context: DecisionContext, criteria: [...], lenses: [...] } })  // omit → 8 criteria / 6 lenses

Gate	Expect
Phases	Generate → Score → Stress-test
Agent labels	gen:<lensKey> per lens, judge:1..3, refute:<optionId> for top ~3
Logs	Generated N options (N = lens count), Scored by 3 judges; stress-testing top 3
Return shape	{ options[], criteria[], judgeCount, stressTested[] }; each option has scores (per-criterion {mean, spread, notes}) and adversarial ({verdict, killerRisks})

Save the return verbatim to build-options-output.json, then:

.claude/skills/fs-flagship-chain-campaign/scripts/gate.sh phase1 research-output.json build-options-output.json

Expected: GATE PASS: phase1 …, exit 0. Then complete the chain-proof by feeding a decision-data.json (assembled per build-options SKILL.md Phase 2) into:

python3 plugins/build-options/assets/build_matrix.py decision-data.json --out <out-dir>

Expected: HTML <out-dir>/<slug>.html <n> chars printed (the script asserts placeholder injection); the rendered file is the measured artifact.

If you see X instead:

• An option's scores is {} / gate fails "non-numeric scores" → judges returned wrong optionId keys; the aggregation silently skips mismatches. Re-run judges or fix ids — do NOT hand-fill scores.
• judgeCount < 3 → some judge agents returned null and were filtered; the panel is weaker. Gate allows ≥2; below that the "independent panel" claim is false — rerun.
• A stress-tested option shows adversarial.verdict: "survive" with empty killerRisks → KNOWN GOTCHA: a skeptic returning a mismatched optionId is silently replaced by a default {verdict:'survive', killerRisks:[]}. An empty killerRisks on a top-3 option is suspicious — check the refute:<id> labels matched real option ids before trusting "survive".

Promotion: when both gates pass, the proving run's evidence (saved JSONs + gate output + rendered HTML) goes into a dated spec via fs-change-control; the SKILL.md "syntax-checked only" admissions in both plugins may then be updated to "proven on (see spec)" — never simply deleted. Measurement artifacts from this run are exactly what fs-research-frontier's measured-not-claimed program wants — save token counts if the host reports them.

Phase 2 — Rewrite prd/tasks into house style (licensed by operator 2026-07-02)

plugins/prd/SKILL.md (200 lines) and plugins/tasks/SKILL.md (480 lines) are imports from an external "compound engineering" methodology: emoji ❌/✅ rubrics, no references/assets, no model routing, no Announce convention, and tasks hard-depends on the third-party agent-browser CLI (22 mentions, all in plugins/tasks/SKILL.md as of 2026-07-02 — an unstated install dependency).

Do the rewrite in place (same plugin dirs — the chain's slash-names and handoff contracts /tasks/prd-*.md → prd.json must not change). Target style lives in fs-skill-style-guide (extracted from the golden four); do NOT copy style from prd/tasks themselves or tufte-viz.

Genericize agent-browser exactly the way Codex became "external executor": the capability becomes a generic "browser executor" with agent-browser named only as an example ("such as agent-browser"), commands expressed as capability patterns (open/snapshot/click/fill/screenshot/ console-check), and the acceptance-criteria table kept — it is the import's genuinely valuable idea (boolean, machine-verifiable criteria).

Measurable acceptance criteria (all must hold before promotion):

#	Check	Command	Expect
1	Frontmatter exactly name + description (folded >-), trigger-rich	head -12 plugins/{prd,tasks}/SKILL.md	matches golden-four pattern
2	No hard third-party dep	grep -rc "agent-browser" plugins/tasks/SKILL.md	only inside "such as" example mentions (target ≤2)
3	No off-style emoji rubric	grep -c "❌|✅" plugins/{prd,tasks}/SKILL.md	0
4	Length in house norm	wc -l plugins/{prd,tasks}/SKILL.md	~78–233 each
5	House closers present	grep for Known limitations, ## References	present
6	Handoff contract unchanged	build-options SKILL.md Phase 4 still true: PRD lands in /tasks/prd-*.md; tasks emits prd.json with passes:false	verified by reading, ideally by one live handoff
7	Nothing else broke	gate.sh phase0	exit 0

Route through fs-change-control (spec naming a before/after for each criterion → plan → operator commits). Per-plugin READMEs (marketing-facing template) follow in the same spec.

Phase 3 — Site truth-up (OPERATOR-GATED — do not start without explicit approval)

docs/index.html violates the repo's own no-oversell doctrine (defect ledger #6, verified still open 2026-07-02): two ~0 tokens cost chips and one "runs itself weekly" claim. The exact strings and line numbers live in fs-failure-archaeology entry 6; the candidate corrections table lives in fs-site-and-positioning — this phase points there rather than duplicating them.

Mechanics and the sibling-site coordination requirement live in fs-site-and-positioning; the push itself goes through fs-release-and-publish (master is live-publishing). Measured success: grep -c "~0 tokens\|runs itself weekly" docs/index.html → 0, and the deployed page reflects it.

Fenced wrong paths (do not go here)

Fence	Why
F1: Do NOT regenerate matrix.template.html from scratch	The rescued copy is canonical; a regenerated one is an untested lookalike that silently changes the rendered artifact. If it goes missing, restore from git history (post-commit) or the plugin cache — never re-author.
F2: No GitHub Actions, ever	Standing org-wide rule since 2026-06-19 (billing-blocked + operator preference). All gates are LOCAL scripts.
F3: No mocked proving run counted as proof	The node:test harness already mocks; the unknown is the live runtime.
F4: Never hand-compute weighted totals	build_matrix.py + weighting.py are the single source of truth; hand math desyncs viz from data.
F5: Don't delete the honesty admissions	"syntax-checked only" lines get updated with evidence, not removed.
F6: Don't edit docs/index.html outside Phase 3's operator gate	Site edits need a coordinated spec (sibling catalog site must stay in sync manually).

Quick reference

# status of the campaign's ground truth
git -C <repo> status --short                       # phase-0 files still uncommitted?
.claude/skills/fs-flagship-chain-campaign/scripts/gate.sh phase0
.claude/skills/fs-flagship-chain-campaign/scripts/gate.sh phase1 research-output.json build-options-output.json
python3 plugins/build-options/assets/build_matrix.py decision-data.json --out out/   # rendered-artifact check

Provenance and maintenance

All facts verified 2026-07-02 against HEAD 2e4c9dd + the uncommitted phase-0 working tree. Re-verify before trusting:

• Phase-0 state: git -C <repo> status --short (expect the seven-entry working tree listed in Phase 0 until the operator commits; after commit, gate.sh phase0 is the only check that matters).
• Test counts (9/6/5): gate.sh phase0 — it embeds all three suites.
• Workflow return shapes: grep -n "^return {" plugins/market-validation/assets/research-workflow.js plugins/build-options/assets/build-options-workflow.js (lines 146 and 129 as of 2026-07-02).
• honesty admissions graduated: grep -rln "syntax-checked only" plugins/*/SKILL.md matches ONLY skill-style-guide (which teaches the convention) since the 2026-07-02 promotion — market-validation and build-options now read "proven on 2026-07-02"; the spec records the evidence.
• agent-browser count: grep -rc "agent-browser" plugins/tasks/SKILL.md (1 since the 2026-07-02 rewrite; >2 = regression).
• Site oversell strings: grep -n "~0 tokens\|runs itself weekly" docs/index.html (0 since 2026-07-02; any hit = regression).
• Gate script self-test: it was verified 2026-07-02 to PASS on real phase-0 state and to FAIL on synthetic bad phase-1 JSON (leaked drop verdict; empty scores). If you edit it, re-run both directions.
• Reference-run cost figures (~1.5M tokens/~38 agents/~50 min): sourced from plugins/market-validation/SKILL.md "Before you start" — update there first, here second.