تشغيل أي مهارة في Manus بنقرة واحدة

feature-pipeline

النجوم٠

التفرعات٠

آخر تحديث٢٢ يونيو ٢٠٢٦ في ١٦:٢٣

Triggers on intent-to-run-the-pipeline phrases. Use when the user says "run the feature pipeline", "start the pipeline for X", "kick off the orchestration", "automate feature X end-to-end", "feature pipeline", "run the pipeline for X", "start the orchestration", or "pipeline a new feature". Does NOT trigger on "create an issue", "review a PR", "open a PR", "merge a branch", or "deploy" — those stay in their own skills.

التثبيت

التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.

تشغيل في Manus

المصدر

julianken

julianken/language-agent

فتح مستودع GitHub عرض مستودعات المنشئ

تنزيل

تشغيل في Manus

المهن ذات الصلةSOC

استنادا إلى تصنيف SOC المهني

مطوّرو البرمجياتمهن الحاسوب والرياضيات·SOC 15-1252

مستكشف الملفات

2 ملفات

SKILL.md

readonly

المزيد من هذا المستودع

نفس المستودع

figma-design

julianken/language-agent

Use when building or modifying Figma frames for a feature before any issues or code are created. Triggers on: - "build the Figma frames for feature X" - "create the WIP design page for X" - "design feature X in Figma" - "lay out the screens in Figma before issue creation" - "build the pre-code design for X" - "create WIP frames for the X feature" - "set up the Figma WIP page for X" - "place the designs for X on the WIP page" Does NOT trigger on: - "review the Figma design" → use `reviewing-figma-designs` (P9) - "approve the Figma frames" → use `reviewing-figma-designs` (P9) - "implement the component" → use the implementer skill - "update DESIGN.md tokens" → DESIGN.md edits are out of scope for this skill - "judge whether the frames are correct" → use `reviewing-figma-designs` (P9) - "does this design pass?" → use `reviewing-figma-designs` (P9) <example> Context: The feature-pipeline engine is at phase 3 (pre-code Figma design) for a feature. user: "Build the Figma frames for feature-slug from the epic spec

2026-07-030

reviewing-figma-designs

julianken/language-agent

Use when reviewing WIP Figma frames for a feature before any issues or code are created — the pre-code design gate. Triggers on: - "review the Figma design for feature X" - "approve the WIP frames for X" - "design review before issue creation" - "gate the pipeline on the Figma design" - "check the Figma frames before we create issues for X" - "run the pre-code Figma review for X" - "does the Figma design for X pass the design system check?" - "review the WIP page frames before coding starts on X" Does NOT trigger on: - "review the PR" → use `reviewing` (PR diff review) - "review the built UI" → use `design-reviewer` (Playwright/built-UI pass) - "build the frames" → use `figma-design` (P8, the frame builder) - "implement the component" → use the implementer skill - "update DESIGN.md tokens" → out of scope for this skill - "judge whether code matches the design" → use `design-reviewer` <example> Context: The feature-pipeline engine is at phase 4 (pre-code Figma review) for a feature. The figma-design (P8) ski

2026-07-030

reviewing

julianken/language-agent

Use when reviewing a PR diff on julianken/language-agent and you need the anti-slop rubric — verify-before-claim, ≤3 findings, severity tiers, mandatory second pass, prompt-injection defense. Bot-agnostic: applies whether a human or any bot identity posts the review. Triggers on "review this diff", "anti-slop review", "review rubric", "what severity is this finding". Self-contained for worktree dispatch; carries no credentials.

2026-07-030

creating-prs

julianken/language-agent

Use when drafting a PR body or deciding what goes in the description on julianken/language-agent. Triggers on "draft a PR description", "what should the PR body say", "fill the PR template", "PR body sections", or a PR that came back with template-shaped feedback. The generic five-section method, conventional-commit rules, and plan-reference discipline — bot-agnostic, no credentials. Self-contained for worktree dispatch.

2026-06-220

epic-authoring

julianken/language-agent

Use when a new feature needs a research-backed epic spec synthesized into tmp/docs/ before child issues are drafted. Triggers on "research and write the epic", "author the epic spec", "write the feature epic", "run the research phase for the epic", "synthesize a feature brief into an epic", "draft the epic from the brainstorm brief", "produce the epic spec for issue fan-out", "build the epic for this feature", "epic research for <feature>", or "I need an epic spec before filing child issues". Invoke when phase(2) of the feature pipeline fires, when the caller has a brainstorm brief and needs a structured epic before calling issue-authoring, or when any session asks to research and produce an epic spec for a new product feature. Self-contained for worktree dispatch — all hard constraints restated in-body.

2026-06-220

issue-authoring

julianken/language-agent

Use when opening or rewriting an implementation issue or plan spec on julianken/language-agent. Triggers on "create issue", "write the issue", "issue spec", "implementation plan issue", or batch prep/planning work. Self-contained for worktree dispatch.

2026-06-220

name

feature-pipeline

description

Feature Pipeline — driving skill

Announce at start: "I'm using the feature-pipeline skill to run the automated feature pipeline."

This is the human-facing entry point for the feature pipeline system. It makes the P5 engine (.claude/workflows/feature-pipeline.js) discoverable and correctly invocable by a fresh agent. It does NOT duplicate the engine logic — it documents the args contract, the front-door brainstorm, and the stage map so a cold-start worktree-isolated agent is self-contained.

Hard constraints (restated for worktree isolation)

These constraints are restated here in full because worktree-isolated agents do NOT load AGENTS.md — this skill is their sole instruction surface.

Repo slug: julianken/language-agent; local folder is language-agent/.
Engine path: .claude/workflows/feature-pipeline.js — a saved Claude Code Workflow script; invoke it via the Workflow tool, NOT as a shell command.
No human pauses mid-run. The engine runs fully autonomously. All review gates are dispatched as fresh-context @julianken-bot subagents. Blockers go to a RESOLUTION sub-workflow (diagnose→fix→verify). The engine escalates a HIL: note only for a genuine info-need or a truly-unresolvable problem (≥2 failed RESOLUTION cycles on the same blocker, or a missing arg only the owner can supply).
RESOLUTION sub-workflow (diagnose→fix→verify): when a phase encounters a blocker (failed gate, build error, bot REQUEST_CHANGES with no actionable fix path), the engine dispatches a resolution sub-workflow: (1) diagnose — read the error/finding, identify root cause; (2) fix — apply the minimal change; (3) verify — confirm the fix satisfies the gate before re-dispatching. Only after ≥2 failed RESOLUTION cycles on the same blocker does the engine emit a HIL: escalation note and halt the affected phase.
Model routing: opus for review gates and synthesis stages; sonnet for mechanical stages (issue fan-out, PR-body fill). A model arg overrides only the mechanical-stage tier; review gates always use opus regardless.
Worktree policy: every change runs in a linked git worktree; never commit on main.
Never commit briefs, plans, or tmp/ content. Plans live in gitignored tmp/docs/, never in docs/plans/ or any committed path.
Bot identity: @julianken-bot posts reviews; the main session never runs gh pr review (that posts as @julianken and fails the per-HEAD branch ruleset).

One-HIL-pause framing

The autonomous run (the engine) has zero mid-run pauses — the engine never blocks for the human. The single HIL touchpoint is the FRONT-DOOR brainstorm (phase 0, below), which runs in the MAIN session BEFORE the engine is invoked. So the model is:

Human-IN-the-loop for the irreversible product decision (front-door brainstorm, main session, phase 0).
Human-ON-the-loop for everything after (watch the dashboard; intervene only via a HIL: comment when the engine escalates a true info-need).

Phase 0 — FRONT-DOOR brainstorm (runs in MAIN session, before the engine)

The Workflow runtime cannot take mid-run input, so the brainstorm cannot run inside feature-pipeline.js. Before invoking the engine, the driving skill runs the one sanctioned HIL touchpoint in the MAIN session:

Invoke superpowers:brainstorming — one question at a time, propose 2–3 approaches, present a design, HARD-GATE on the user's approval before any implementation.
On approval, write the brief to the GITIGNORED path tmp/docs/<featureSlug>-brief.md — NOT docs/superpowers/specs/ (the brainstorming skill's committed-spec default). This is a one-line documented INSTANCE ADAPTATION: superpowers:brainstorming explicitly honors "User preferences for spec location override this default", and we do NOT invoke its terminal writing-plans step (epic-authoring is our downstream). The brief is written to the gitignored tmp/docs/ path; it is never committed.
Pass brainstormBriefPath (the path to the brief) as a start arg to the engine. The brief is immutable start input the engine only READS — the standard "no mid-run Update primitive ⇒ externalize the human decision as a start arg" resolution.

Brief required-field shape

The brief at tmp/docs/<featureSlug>-brief.md must contain these fields (fixed-field, literal-minded-parseable):

PROBLEM — one-paragraph user-facing goal.
SCOPE-IN — explicit bullet list of what is in scope (the YAGNI boundary).
SCOPE-OUT — explicit bullet list of what is explicitly out of scope.
SUCCESS CRITERIA — observable/testable outcomes that seed the epic's acceptance criteria.
CONSTRAINTS — design/token/perf/a11y/platform constraints, pointing at DESIGN.md where relevant.
uiFeature: — on its OWN greppable line, exactly uiFeature: yes or uiFeature: no plus a one-line justification. Matched with grep -F, no NLP. This is load-bearing: the engine reads it at phase 1 to set the uiFeature flag.

Example brief fragment:

PROBLEM
One paragraph describing the user-facing goal.

SCOPE-IN
- Item 1
- Item 2

SCOPE-OUT
- Item A
- Item B

SUCCESS CRITERIA
- Observable outcome 1
- Observable outcome 2

CONSTRAINTS
- See DESIGN.md §0 for token values
- a11y: WCAG 2.1 AA

uiFeature: yes  <!-- adds visible UI to the note map -->

Args contract

The engine requires Claude Code v2.1.154+ for schema-driven arg prompting (automatic prompting for missing required args). If your version is older, pass all args explicitly in the Workflow invocation — the engine will not prompt you.

Eight args total (five required, three optional):

Arg	Type	Required	Default	Description
`featureSlug`	string	required	—	Kebab-case identifier for the feature, e.g. `pipeline-infra`. Keys stage cards in `pipeline.json`.
`epicSpecPath`	string	required	—	Absolute path to the gitignored `tmp/docs/` epic spec the epic-authoring phase will use (or has used).
`childSpecPaths`	string[]	required	—	Ordered list of absolute `tmp/docs/` spec paths for child issues, dependency-ordered. Children are created first; epic is created last (so the epic can link finalized child numbers).
`repoSlug`	string	required	—	GitHub `owner/repo` slug. For this instance: `julianken/language-agent`.
`brainstormBriefPath`	string	required	—	Absolute path to `tmp/docs/<featureSlug>-brief.md` produced by the front-door brainstorm (phase 0). NOT a prompt-fill arg — the schema-driven missing-arg prompt EXCLUDES `brainstormBriefPath`; the front-door brainstorm is its sole producer. Never ask the user for a path; run the brainstorm to produce it.
`uiFeature`	boolean	optional	derived from brief	Explicit override. The engine DERIVES it at phase 1 from the brief's `uiFeature: yes\|no` marker; this arg forces it. Drives the conditional Figma design+review phases (3 & 4) and the Figma-Design-Verdict gate before fan-out.
`model`	string	optional	per-phase routing	Override model tier. Accepts `'opus'` \| `'sonnet'` \| a full model id. When omitted, the engine applies per-phase routing (opus for review/synthesis stages, sonnet for mechanical stages); a supplied value overrides only the mechanical-stage tier — review gates always use opus.
`dryRun`	boolean	optional	`false`	When `true`, logs each phase decision without mutating GitHub state (no issues created, no PRs opened, no reviews posted, no merges triggered). Use for rehearsal or testing the pipeline wiring without side-effects.

brainstormBriefPath is NOT a prompt-fill arg. The schema-driven missing-arg prompt excludes it. The only way to produce it is the front-door brainstorm (phase 0, above). Never satisfy it by asking the user for a path.

Stage map (twelve phases, zero engine pauses)

Phase 0 is the FRONT-DOOR (main session, before engine invocation). Phases 1–11 are the engine's autonomous run. Phases 3 and 4 are UI-only (skipped when uiFeature == no).

Phase	Title	What it does	Leaf worker	Gate
0	Brainstorm (FRONT-DOOR / HIL)	The one sanctioned human touchpoint, in the MAIN session BEFORE the engine: one question at a time, propose 2–3 approaches, present a design, HARD-GATE on approval; write the brief to gitignored `tmp/docs/<featureSlug>-brief.md`; pass `brainstormBriefPath`	`superpowers:brainstorming` (front-door, not an engine leaf)	HARD-GATE on user approval (inside the brainstorming skill); the brief file existing IS the artifact the engine validates
1	Validate	Brief-validate (`brainstormBriefPath` exists + parses + carries the `uiFeature:` marker; set the `uiFeature` flag; absent ⇒ `AGENT: brainstorm needed` + halt, never conduct the brainstorm) and `project-bootstrap` validate-mode (slug, skills, agents present; `pipeline.json` + `feature-pipeline.schema.json` on main)	`.claude/skills/project-bootstrap/SKILL.md` (validate mode) + engine-local brief check; model: sonnet	None — abort/escalate on failure
2	Epic authoring	CONSUMES the brief; deep-research synthesis → `tmp/docs/` epic spec; carries `uiFeature` into the spec	`.claude/skills/epic-authoring/SKILL.md`; model: opus	None — produces the draft spec
3	Figma design (UI-only)	Build/modify the feature's WIP-page frames from the epic spec's SCOPE-IN + CONSTRAINTS, placing DESIGN.md §0 token values, idempotently	`.claude/skills/figma-design/SKILL.md`; model: opus	None — produces the WIP frames (skipped if `uiFeature == no`)
4	Figma design review (UI-only)	`@julianken-bot` loaded with `reviewing-figma-designs` reviews the WIP frames; writes `Figma-Design-Verdict` to the verdict file; REQUEST_CHANGES loops back to phase 3	`.claude/skills/reviewing-figma-designs/SKILL.md`; `.claude/agents/julianken-bot.md`; model: opus	`Figma-Design-Verdict: APPROVE` in the verdict file (skipped/auto-satisfied if `uiFeature == no`)
5	Issue fan-out	Creates child issues first, then the epic last (so the epic links finalized child numbers); UI children reference the approved WIP frames by node-id	`.claude/skills/issue-authoring/SKILL.md`; model: sonnet	ENTRY precondition (UI features): the Figma verdict file reads APPROVE before any issue is created
6	Plan-review gate	`@julianken-bot` reviews each child issue, then epic LAST; loops REQUEST_CHANGES→fix→re-dispatch until APPROVED vs comment thread	`.claude/skills/issue-plan-review/SKILL.md`; `.claude/agents/julianken-bot.md`; model: opus	All issues APPROVED (epic last)
7	Implement	Parallel implementer dispatch (dependency-ordered); each implementer works in its own linked worktree on a feature branch	Subagent per child issue	None — parallel agents
8	Create PRs	Each implementer opens a PR filling all five template sections; UI PRs attach screenshots before review	`.claude/skills/creating-prs/SKILL.md`; `pr-screenshots-via-user-attachments`; model: sonnet	None — PR open
9	Review loop	`@julianken-bot` reviews each PR (model: opus); parallel with design-reviewer on UI PRs; loops REQUEST_CHANGES→fix→push→re-dispatch-on-new-HEAD until APPROVED vs HEAD	`.claude/skills/reviewing/SKILL.md`; `.claude/agents/julianken-bot.md`; `.claude/agents/design-reviewer.md` (UI PRs); model: opus for gates	`reviewDecision == APPROVED` per current HEAD (polled via gh)
10	Merge	Posts `@Mergifyio queue` standalone comment once APPROVED vs HEAD; Mergify squash-merges	`.claude/skills/pr-workflow/SKILL.md`	`mergedAt` set (polled via gh)
11	Deploy	`deploy.yml` fires on push to `main`; workflow polls until green	gh `statusCheckRollup`	All checks green

Epic-last gate

Children before epic — always. This invariant is enforced as a machine-readable gate in the engine AND stated here for human and cold-start-agent comprehension:

Epic is created AFTER all children — so the epic can link finalized child issue numbers (children are created at phase 5, epic is the final item in that same phase).
Epic plan review runs LAST — after all children are APPROVED by @julianken-bot (phase 6).
Epic is closed LAST — after all child PRs have mergedAt set (phase 10 completes for all children before the epic issue is closed).

No phase 5–10 action targets the epic until the children it describes are already stable.

Resolution-not-HIL policy (restated for worktree isolation)

When a phase encounters a blocker (a failed gate, a build error, a bot REQUEST_CHANGES with no actionable fix path), the engine dispatches a RESOLUTION sub-workflow:

Diagnose — read the error/finding, identify root cause.
Fix — apply the minimal change.
Verify — confirm the fix satisfies the gate before re-dispatching.

Only after ≥2 failed RESOLUTION cycles on the same blocker, or when the blocker is a genuine info-need (e.g. a missing arg only the owner can supply), does the engine emit a HIL: escalation note and halt the affected phase.

The human-ON-the-loop posture: watch the dashboard (tools/pipeline-dashboard/pipeline.json); intervene only via a HIL: comment when the engine escalates. Do not monitor progress by re-invoking the engine.

Trigger testing

Should trigger this skill (≥8 examples)

"run the feature pipeline"
"start the pipeline for the tuner"
"kick off the orchestration for scale labels"
"automate the horizontal map feature end-to-end"
"feature pipeline"
"run the pipeline for the ref-overlay feature"
"start the orchestration for the mobile redesign"
"pipeline a new feature"
"use the feature pipeline to ship the settings panel"
"invoke the pipeline workflow for dark mode"

Should NOT trigger this skill (≥8 examples)

"create an issue for the tuner feature" → use .claude/skills/issue-authoring/SKILL.md
"review this PR" → use .claude/skills/reviewing/SKILL.md
"open a PR for branch feat/x" → use .claude/skills/creating-prs/SKILL.md
"merge branch feat/x" → use .claude/skills/pr-workflow/SKILL.md
"deploy to production" → use deploy.yml / infra
"what are the open issues?" → use gh issue list
"review the figma designs for the tuner" → use .claude/skills/reviewing-figma-designs/SKILL.md
"run the brainstorm for scale labels" → use superpowers:brainstorming directly (phase 0 only)
"bootstrap the project" → use .claude/skills/project-bootstrap/SKILL.md
"author the epic for S14" → use .claude/skills/epic-authoring/SKILL.md

References

Engine: .claude/workflows/feature-pipeline.js — the saved Workflow script this skill invokes.
Stage taxonomy: tools/pipeline-dashboard/pipeline.json — the machine-readable source of truth for stage names, phase ids, and status values.
Schema: .claude/workflows/schemas/feature-pipeline.schema.json — the JSON Schema for the args contract.
Review rubric: .claude/skills/reviewing/SKILL.md — bot-agnostic anti-slop rubric used by all @julianken-bot review gates.
Issue authoring: .claude/skills/issue-authoring/SKILL.md — shape for implementation issues created at phase 5.
Plan review: .claude/skills/issue-plan-review/SKILL.md — gate rubric used at phase 6.
PR method: .claude/skills/creating-prs/SKILL.md — five-section PR body discipline used at phase 8.
Instance facts + merge: .claude/skills/pr-workflow/SKILL.md — the @Mergifyio queue trigger and ruleset facts.

No-drift mirror note

This skill is NOT one of the paired no-drift skill copies tracked in AGENTS.md § Skill ownership (those are creating-prs ↔ user-level creating-prs and reviewing ↔ user-level reviewing-as-julianken-bot). A user-level feature-pipeline namesake does not exist at time of authoring. If one is created later, the no-drift rule applies from that point: changes to either copy must update the other in the same PR, and the PR Summary must say so.