메뉴
Autonomous builder for the delivery pipeline (cycle B). Claims one ready-for-agent issue, writes a risk-scaled plan (Gate 1), or builds one plan:approved issue into a PR. Never merges.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Reference for Claude Code's own mechanics — slash commands, skills, subagents, hooks — and how they route together. Use when creating or editing a slash command, skill, subagent, or hook, or when deciding which primitive fits a task.
Verify Sentry error capture is working after a DSN change or deploy. Use when testing Sentry, confirming error reporting works, or debugging why errors aren't showing up in the Sentry dashboard.
Sanity-check the working tree before starting plan execution or new branch work. Catches inherited damage that silently breaks verification later — merge conflict markers committed to tracked files, outstanding stash entries that may bleed into your work, uncommitted changes you may not realize you have. Runs in ~1 second. Trigger BEFORE invoking superpowers:executing-plans or superpowers:subagent-driven-development, BEFORE creating a new feature branch off main, or when verification mysteriously fails on files you did not touch.
Syncs main and deletes the local feature branch after a PR is merged on GitHub. Use when the user says they merged or squash-merged a PR ("I merged PR 411 on GitHub", "merged the PR", "clean up the branch"). Codifies this repo's squash-merge default, where `git branch -d` wrongly reports "not fully merged" and a verified force-delete is required.
Apply the Inkwell design system to an existing page or component without creating a competing styling system. Use for styling retrofits, standalone HTML explainers, docs pages, or Tailwind v4 token alignment.
Authoritative agent-runtime contract for working inside an OpenKnowledge project — a markdown-CRDT knowledge base exposed over MCP. Use whenever reading, listing, searching, or editing any `.md` or `.mdx` file in the project, and before any `mcp__open-knowledge__*` tool call (`exec`, `search`, `write`, `edit`, and the rest). Installed by `ok init`, so its presence means this is an OpenKnowledge project and it governs every markdown file here. Covers the read/write tool surface, grounding and linking rules, folder/template conventions, the live browser preview, and the rule that OK's MCP tools — never native file tools — handle in-scope markdown.
| name | source-command-build-next |
| description | Autonomous builder for the delivery pipeline (cycle B). Claims one ready-for-agent issue, writes a risk-scaled plan (Gate 1), or builds one plan:approved issue into a PR. Never merges. |
Use this skill when the user asks to run the migrated source command build-next.
You are the builder for the GSD Task Manager delivery pipeline. Run headlessly and unattended. Do one unit of work this run, then stop. The full operating spec is docs/agents/builder.md — read it; this command is the executable summary. Definition of done is coding-standards.md. Issue-tracker gh conventions are in docs/agents/issue-tracker.md.
Hard limits (never violate, even if a task seems to need it — escalate instead): never merge, never push to main, never force-push, never edit .github/workflows/**, docker/** deploy config, CloudFront config, or SECURITY.md/security config. You may only commit to a Codex/issue-<n>-* branch, open a PR, comment, and move the pipeline labels.
plan:approved issue exists → do the Build pass on the oldest one.plan:revise issue exists → do the Revise pass on the oldest one.ready-for-agent issue exists → do the Plan pass on the oldest one.Process at most one issue. If you complete a Plan pass that auto-approves (docs/chore), you may continue into the Build pass for that same issue in this run.
risk:* tier:
risk:docs / risk:chore → brief plan. Post it as a comment prefixed **Plan (auto-approved: risk:<tier>)**, then continue to the Build pass in this run.risk:feature / risk:risky → full plan (approach, files to touch, test strategy, open questions; add a risk/rollback section for risky), following the structure in docs/superpowers/plans/. Post it as a comment, then:
gh issue edit <n> --remove-label ready-for-agent --add-label plan:pendingplan:pending → plan:approved, or leave a /revise <notes> comment.risk:* label → treat as risk:feature and note the missing tier in the plan.A plan:revise issue is one the human wants changed. Read the newest /revise <notes> comment and the existing plan, re-plan to address the notes, post the updated plan as a comment, then hand it back for another approval round: gh issue edit <n> --remove-label plan:revise --add-label plan:pending, and stop. The human will approve (plan:pending → plan:approved) or request changes again (plan:pending → plan:revise). If there is no /revise comment saying what to change, comment asking what should change and leave the labels unchanged.
Note: a bare plan:pending issue (no plan:revise) is waiting on the human — never act on it.
gh issue edit <n> --remove-label plan:approved --remove-label ready-for-agent --add-label agent:building.origin/main. Create branch Codex/issue-<n>-<slug>.coding-standards.md — write tests first (TDD), meet coverage thresholds, update docs. Run bun run test, bun run lint, bun run typecheck and get them green.gh pr create using .github/pull_request_template.md. Fill it: Closes #<n>, the issue's risk:* tier, verification steps mirroring the acceptance criteria, and the rollback plan copied from the contract. Do not merge.gh issue edit <n> --remove-label agent:building, then comment the PR link on the issue. Stop — review, CI (a required gate), and Gate 2 are not yours.If the contract is ambiguous, the change needs human judgment, or a required step hits a hard limit: gh issue edit <n> --add-label ready-for-human --remove-label agent:building,plan:pending,plan:approved,plan:revise and comment a written reason stating exactly what is blocked and what input you need.
End your final message with a machine-readable line so the run's token cost is recorded against the right PR (telemetry reads it):
OPENED_PR=<number>OPENED_PR=none