| name | yolo |
| description | Take a bug, ticket, or feature from investigation all the way to a draft PR in one autonomous run — `/goal` planning plus the full ystack chain (/build → /go → /qa → /review → /docs → /pr), orchestrated end-to-end without stopping between phases. Use this skill when the user says 'yolo', '/yolo', 'ship it', 'just ship it', 'do the whole thing', 'full pipeline', 'investigate and fix and PR', 'end to end', 'take this to a PR', or hands over a ticket or bug and wants it shipped autonomously rather than only planned. If the user wants ONLY a plan, that's /goal or /build; for a trivial change, that's /quick; to handle reviewer comments on an existing PR, that's /address-review.
|
| compatibility | Designed for Claude Code |
| metadata | {"user-invocable":"true"} |
/yolo — Investigate → Ship, Autonomously
You run the entire ystack workflow in one pass: investigate the request, plan it,
build it, verify it, review it, document it, and open a draft pull request — without
stopping between phases. The value is a clean hand-off: the user gives you a ticket or a
bug and gets back a reviewable draft PR, not a pile of half-finished steps.
You orchestrate; you do not re-implement. Each phase defers to its command — Claude
Code's /goal for planning, then the ystack skills (/build, /go, /qa, /review,
/docs, /pr). Follow each one's own spec — /yolo only chains them and keeps moving.
Proceed autonomously. Don't ask for confirmation between phases. Surface only the
decisions that genuinely need a human (see Guardrails). When you finish, stop at the draft
PR and summarize.
When NOT to use
- The user wants only a plan → use
/goal or /build and stop.
- The change is a trivial fix / chore (typo, bump, config) → use
/quick.
- The user wants to address reviewer comments on an existing PR → use
/address-review.
Phase 1 — Investigate (evidence-first)
Establish the real requirement and root cause before touching code. First-pass
misdiagnosis is the most common way an autonomous run goes wrong.
- If given a ticket reference, pull it first (e.g. the project's issue tracker / MCP)
for the actual requirements — don't work from the one-line summary.
- Reproduce the bug or pin down the feature's scope.
- Trace it through the code: confirm the true handler, data path, and boundary.
Don't guess the event, column, or function — verify against the code and git history.
- Keep all investigation read-only (read-only DB queries,
git log/git blame). No
writes until the plan exists.
Phase 2 — Plan (/goal → /build)
Lead with Claude Code's /goal to capture the goal, the approach, and the trade-offs,
then run ystack's /build to turn it into a module-scoped plan. /build writes
.context/<feature>/PLAN.md and DECISIONS.md with goal-backward success criteria and
an explicit no-over-engineering stance (inline vs. helper, where the source of truth
lives).
Proceed once the plan is written. Pause and ask only if the design is genuinely
ambiguous (two reasonable architectures with materially different blast radius).
Phase 3 — Build (/go)
Run /go. Implement the plan as atomic, scoped commits, one per task. Follow the
plan; if reality deviates, follow /go's deviation rules rather than silently changing
the approach.
Phase 4 — Verify (/qa)
Run /qa. Execute the repo's real CI commands (typecheck / lint / tests, as /qa
detects them) and any plan-driven checks, then fix anything the change introduced until
the suite is green. This gate is non-negotiable: never carry a red or unknown test
state into the next phase. A green-looking phase hiding a broken test is exactly the
failure to prevent.
Phase 5 — Self-review (/review)
Run /review on your own diff. Address real P0/P1 findings, verified against the code —
don't invent findings to look thorough, and don't wave away a real one. Re-verify after
any fix.
Phase 6 — Docs (/docs)
Run /docs. Update every doc page affected by the completed work so docs and reality
stay aligned. Never document planned or in-progress work — ystack's hard rule.
Phase 7 — PR (/pr, draft)
Run /pr. Open a draft PR off a feature branch targeting the repo's default branch
(resolved dynamically, e.g. git symbolic-ref refs/remotes/origin/HEAD). If the user
passed --base <branch> to /yolo, forward it as /pr --base <branch> so the PR targets
that branch instead (e.g. --base staging). Write a clear English title (Conventional
Commits) and body (summary, changes, verification, test plan), and proactively answer the
review questions you can predict.
Phase 8 — Hand back
Stop. Summarize what shipped — feature, commits, files changed, verification status, PR
URL — and surface only the decisions that still need a human.
Guardrails
- Branch off and target the default branch (or the
--base <branch> you were given).
Never commit to or push the base / release branch directly. /yolo ends at a draft
PR — never a merge.
- Atomic commits, scoped per task.
- Never mark a phase done with failing tests or typecheck. The
/qa gate is hard.
- Read-only data access unless the user explicitly authorizes writes.
- Hand back on credentialed or external actions — payment/billing, production data,
deploys, anything needing secrets you don't have. Don't try to power through an
auth/env failure; report it and stop. That's where autonomous runs stall.
- Stop and ask before anything destructive or outward-facing beyond opening the draft
PR itself.
What This Skill Does NOT Do
- Does not invent its own planning, build, QA, review, doc, or PR logic — it runs the
existing skills and follows their specs.
- Does not merge or publish. Only opens a draft PR; merging is a human decision.
- Does not force-push, and does not touch the default or a release branch.
- Does not replace
/quick for trivial changes or /address-review for reviewer
feedback.
