| name | feature-workflow |
| description | The umbrella dev loop for any non-trivial feature, fix, or refactor in this repo —
invoke at TASK START, before exploring or editing. It does not add new rules; it
sequences the validated discipline skills at the phase boundaries where the recorded
Opus sessions put them in the wrong order (gates at the end instead of per phase,
review after merge instead of before done, decisions locked before the user saw
them). Routes to: decision-gate → probe-state-before-acting → implement with
per-phase verification → verify-before-done → grill-your-own-work → /ship.
|
| allowed-tools | ["Skill","Read","Bash"] |
| metadata | {"author":"distilled-from-fable-5-sessions","version":"0.1.0-composition"} |
Feature Workflow
Why a router: the evidence shows Opus has every individual discipline — it runs
gates, reviews, plans well — but sequences them wrong. The Circle port ran ALL its
verification once, at the end (932ee514 L945 → 9 bugs found post-"done"); the
meal-card sessions reviewed after shipping (8c0bed51, 3-surface revert); the barcode
session locked a schema design before the user saw it (8013b7ec L66). Right pieces,
wrong order. This skill is the order.
Phase 0 — Orient (before ANY edit)
- AGENTS.md §2.2 preflight: invoke the matching stack skills (vercel-react-best-practices
/ nham-design / mobile docs) — the most-skipped mandatory step on record.
- Load probe-state-before-acting: fetch, check whether the branch/work already
exists, confirm the worktree root. (Worktree from latest main per repo convention.)
- Restate the ask. If it's multi-part feedback, produce a numbered parse of every
distinct ask first, and close each explicitly in your wrap. (A dedicated
absorb-steering skill FAILED blind validation — its deferral framing correlated with
under-delivery — so this router keeps only the parse/close mechanic, which tied, and
drops the rest.)
Phase 1 — Decide (before building)
Load decision-gate. Split forks from implementation. Surface forks as ONE
decision-table question (or the top of your report if non-interactive). Schema/auth/
export-semantics/UX-paradigm forks are always gated. Only then plan the build — plan
depth itself is your call; the evidence shows no planning-quality gap, so don't
over-ceremonialize small tasks.
Phase 2 — Implement (verification per phase, not terminal)
- After each coherent chunk: targeted tests + analyze/tsc on the touched area. Do NOT
save all verification for the end — terminal-only verification is the single
costliest recorded pattern.
- Unfamiliar library → read
node_modules/**/*.d.ts + Context7 before coding against it.
- Anything unexpected during your own QA → load root-cause-first before continuing.
- Delegating? Load delegate-and-verify first (REPORT-ONLY reviewers, prescribed fix
designs, never relay self-reports).
Phase 3 — Close (before saying "done")
- Load verify-before-done: full gates from the right root + behavior-altitude
observation + evidence-audited report.
- For features beyond a trivial patch, load grill-your-own-work: adversarial pass
on your own diff before the user sees it. This is the step the user used to trigger
manually by switching models — make it automatic.
- Ship via /ship only when the user asks; branch naming + conventional commits per
AGENTS.md.
Session-long invariants
- Steering arrives → parse it into numbered asks, act, close item-by-item in the wrap.
- Context pressure rising → verify FIRST, then summarize; externalize state (plan file,
tasks, memory) instead of holding it in context.
- End of session → AGENTS.md retrospective + memory write for durable lessons.
Status: COMPOSITION — not independently validated. The four routed discipline
skills were blind-validated 8-0 individually (batch 1) and the routed batch-2 skills
carry their own validation records; this router itself is ordering logic distilled from
the recorded sequencing failures, and a router cannot be trap-tested in a single-shot
arm. Treat its value claim as: "the pieces are proven; the order is evidence-derived
but unproven as a bundle."
Evidence: .claude/skills/_evidence/findings.md; validation:
.claude/skills/_evidence/validation/results.md.
Last verified: 2026-07-09. Drift re-check: ls .claude/skills/{verify-before-done,grill-your-own-work,root-cause-first,probe-state-before-acting,decision-gate,delegate-and-verify}/SKILL.md
