feat(skills): /feature — idea → create-prd → create-stories, gated (#208)

[thomasluizon]

What

Adds the /feature orchestrator skill at .claude/skills/feature/SKILL.md — a thin chain that drives a rough idea through the full backlog pipeline so it becomes well-formed GitHub issues without hand-running each step.

Flow (gated)

/feature "<idea>"
  → Stage 1: grill (grill-me posture; skipped on --cold)
  → Stage 2: /create-prd <kebab>.prd.md   (or /prd-interactive on --cold / empty idea)
  → GATE 1 (HARD, owned here): confirm PRD — yes / edit <notes> / abort
  → Stage 3: /create-stories .claude/PRDs/<name>.prd.md   (forwards --milestone / --no-create)
  → GATE 2 (HARD, REUSED): /create-stories' own Phase 6.1 "About to create N issues … Confirm?"
  → Report (suggests, never runs, the per-issue /prime→/plan→/implement loop)

• Thin orchestrator — references the shipped flows by invocation name (grill-me, /create-prd, /prd-interactive, /create-stories). No PRD or story logic is reimplemented; the chained skills own the templates, labels, and issue bodies.
• GATE 1 is the only gate this skill owns (PRD→stories boundary). GATE 2 reuses /create-stories Phase 6.1 verbatim — not re-spelled, not auto-answered. --no-create skips issue creation entirely.
• Usage: /feature "<idea>" [--cold] [--milestone "MVP"] [--no-create].

Also

One-line driven-shortcut pointer added to WORKFLOW.md under the medium/large path; the manual steps stay documented.

Validation

Markdown skill — no npm/dotnet. Static read-through confirms: frontmatter name: feature + description, usage line + all flags, both gates present, Stages 2–3 invoke the real flows by name, zero .claude/commands/ path references. Manual dry-run recommended (warm --no-create, gate-abort, edit-loop, cold-start).

Dependencies / merge note

Depends on #206 (the command → skill migration); the chained flows are invoked by their stable invocation names, which survive that migration. Touches WORKFLOW.md — may conflict with #206's WORKFLOW.md edit; resolve at merge. Merge after #206.

Closes #208

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
orbit-ui-mobile-web	Ignored		Jun 19, 2026 8:28pm

[claude]

This lands cleanly. The feature skill is a well-structured thin orchestrator — it chains the existing flows by invocation name rather than reimplementing them, owns exactly one gate (PRD-to-stories), and correctly reuses create-stories Phase 6.1 gate rather than duplicating it. The routing table (warm/cold), the gate branches (yes/edit/abort), and the guardrails section are all correct and coherent. The one-line addition to WORKFLOW.md is a good pointer without cluttering the manual steps. No hard-rule violations, no dead references in code, no parity concerns (tooling changes only, not app code). Good to go.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[claude]

Code Review: PR #238

Scope: PR #238 — feat(skills): /feature — idea → create-prd → create-stories, gated (#208)
Recommendation: NEEDS WORK

Summary

This PR adds a new /feature orchestrator skill (.claude/skills/feature/SKILL.md) and a two-line pointer in WORKFLOW.md. The overall design is sound — thin orchestrator, no reimplemented logic, two hard gates — but two issues need resolution before merge: one High (a hardcoded Windows-absolute path to grill-me that will silently fail on CI and other machines), and one Medium (Stage 3 prose that elides the --no-create conditionality on Phase 6). Two Low/Info observations are also noted.

---

Findings

Critical

None.

High

[HIGH] Stage 1 hardcodes a Windows-absolute path to grill-me that is non-portable and will break on CI/Linux

• dimension: Correctness (#1)
• location: .claude/skills/feature/SKILL.md:37
• issue: Stage 1 instructs: "Apply the grill-me posture (C:/Users/thoma/.claude/skills/grill-me/SKILL.md)". The sibling /execute skill references the same skill with no filesystem path — the harness resolves skills by name. Hardcoding C:/Users/thoma/... makes this path meaningless on the Linux CI runner and on any collaborator's machine, and leaks a local developer path into a committed file.
• risk: Any agent reading this file literally will encounter a dead path on CI or any non-developer machine. Even for the primary developer it is fragile across machines or user-directory renames.
• fix: Replace the parenthetical with the invocation-name-only reference used everywhere else: "Apply the grill-me posture" — no path, no parenthetical. The harness resolves the skill by name; filesystem paths belong in .mcp.json / settings, not a SKILL.md.
• reference: CLAUDE.md rule 1 (root-cause, no workarounds); /execute SKILL.md lines 50-53 (established pattern)

---

Medium

[MEDIUM] Stage 3 prose implies Phase 6 of /create-stories always fires — omits the --no-create conditionality

• dimension: Correctness (#1)
• location: .claude/skills/feature/SKILL.md:72-74
• issue: Stage 3 says: "It writes .claude/stories/<name>.md (its Phase 5) and then proceeds to its own Phase 6." This reads as unconditional, but /create-stories Phase 6 opens with: "Skip this phase if --no-create was passed." The flag is documented in the inputs table but Stage 3's prose does not reflect it — an agent reading Stage 3 in isolation would not know Phase 6 is conditional.
• risk: Low runtime risk (the flag is documented elsewhere in the file), but the description could cause an agent to mis-describe the flow to the user or miscount the gates in the --no-create path.
• fix: Change Stage 3 to: "It writes .claude/stories/<name>.md (Phase 5). If --no-create was NOT passed, it then proceeds to its own Phase 6 (issue creation gate)." This matches /create-stories exactly.
• reference: CLAUDE.md rule 1 (accuracy of chained-flow descriptions)

---

Low / Info

[LOW] Stage 2 cold-path aside is a load-bearing instruction buried in a parenthetical

• dimension: Comment policy (#4, informational — linter does not cover .md files)
• location: .claude/skills/feature/SKILL.md:58-59
• issue: "read the path back from the SUMMARY before Stage 3 rather than assuming it" is a correctness-critical mechanical instruction easy to miss in prose. Not a lint violation.
• risk: None immediate.
• fix: Optional — extract as a titled note block for scanability.

[INFO] WORKFLOW.md two-line addition is accurate, minimal, and well-placed — no findings.

---

Subagents

Agent	Verdict
parity-checker	N/A
i18n-syncer	N/A
contract-aligner	N/A
security-reviewer	N/A

Validation

Check	Result
Lint	N/A — Markdown skill file
Type check	N/A
Tests	N/A
Build (api)	N/A

What's good

• Thin orchestrator discipline upheld. Invokes grill-me, /create-prd, /prd-interactive, /create-stories by invocation name only — exactly the pattern /execute sets.
• GATE 2 reuse is explicitly guarded. "Do NOT auto-answer it, pre-confirm it, or suppress it" prevents the outer-orchestrator failure mode of silently passing the inner skill's gate.
• Routing table is clean and exhaustive. The --cold / non-empty-idea / empty-idea branching in Stage 0 has no overlap or gap.
• Guardrails section is thorough. Covers all the shortcut failure modes: re-implementing PRD logic, double-grilling, running /plan//implement, etc.
• --no-create dry-run path is explicit in both the flag table and GATE 2.

Recommendation

Two one-line prose fixes before merging:

1. Remove the hardcoded Windows path in Stage 1 line 37 — change to simply: "Apply the grill-me posture" (no filesystem path).
2. Tighten Stage 3 to reflect the --no-create conditionality on Phase 6: "If --no-create was NOT passed, it then proceeds to its own Phase 6."

Once those land, the skill is ready to merge.

---

Reviewed by /pr-review against rubric.md