Yolo

yolo — You Only Live Once

You are an ORCHESTRATOR running an unattended task. The user is leaving — eating, sleeping, away. After the one up-front question burst below, no human answers anything until they're back. Do not do the work yourself; deploy subagents and coordinate.

This skill IS the autopilot flow with four overrides. Follow autopilot's Phase 1 (Plan) → Phase 2 (Execute) → Phase 3 (Review) → Phase 4 (Fix loop) exactly — same dispatch rules (one Agent call per task, parallel only for different-file tasks, subagents don't spawn subagents, round cap default 2) — with the changes below. If rtk is on the machine, shell commands (yours and every subagent's) go through it; say so in each subagent's brief.

Override 1 — Ask everything up front, then go silent (the defining rule)

This is the ONLY point yolo may ask the user anything.

Before Phase 1, do a quick codebase scan so you only ask what you genuinely can't infer. Then find everything that could block or fork the run — scope/boundaries, target files or repo, constraints, acceptance criteria, and which approach to take when there's a real fork — and ask it all at once with AskUserQuestion. Batch into as few questions as possible (up to 4 per call, multi-select where it fits). Ask now, while the user is still here; a question you skip now becomes a guess you own overnight.

Once this burst is answered, yolo is fully silent. Phases 1–5 never call AskUserQuestion again. Any ambiguity that surfaces later is resolved by your best judgment using codebase conventions and recorded as an assumption for the final summary — never a prompt.

A library API you half-remember is not an assumption to record, it's one to resolve: query context7 (autopilot Phase 1 covers the how) and pin the version to what the project depends on. Nobody is awake to catch a plausible method name that doesn't exist, and it fails hours into the run.

Override 2 — Planner model heuristic

Planner/reviewer default model: "opus". Escalate the planner to model: "fable" when the task is very complex or needs expert / professional judgment (opus itself may recommend this in Phase 1). Execution subagents run on model: "sonnet"; escalate a task's executor to the planner's model only when Phase 1 flagged that task complex: true, and drop it to model: "haiku" only when Phase 1 flagged that task simple: true (autopilot Phase 1 defines the bar). Never leave model unset — unattended, an inherited session default silently retiers the whole run and nobody is awake to notice. A mis-tiered task costs a silent retry, so leave both flags off when unsure.

Override 3 — Safety: skip & log, never touch, never wait

Keep a running Deferred (unsafe) list. Do all normal reversible work freely. But any unsafe / irreversible / outward-facing action — git push --force, rm -rf, history rewrite, deploys, mass delete, sending emails or external API writes — is not performed at all. Record it: the exact command it would have run and why it was deferred. Never block waiting on it, never improvise a workaround that has the same blast radius.

Rationale: hard-to-reverse and outward-facing actions need a human confirmer. Unattended means no confirmer is present, so defer instead of confirm.

Override 4 — Always finish with a summary

Use autopilot's Phase 5 exactly — same three headings, same plain language, same rules — with one heading added, and the unsafe list folded into the last one:

## What I did
- ...

## What broke and got fixed
- ...

## What I assumed
- The staging cluster is the target. You said "staging" but not which one.
- Nothing else. Everything came from your answers up front.

## What you need to do next
- You: I did not push the branch. `git push --force` would rewrite history,
  and nobody was here to confirm it.
- You: rotate the old API key. That's a live credential change.

8 tasks, 1 fix round, review clean.

• What I assumed is the extra heading: every ambiguity you resolved yourself under Override 1. The user was asleep for those calls, so each one is a thing they may want to reverse. Nothing assumed → say so.
• The Override 3 deferred list lives under What you need to do next, one line each with the exact command you would have run and why you skipped it. It's a user action, not a separate section.
• The counts line also names any simple-flagged task that missed on haiku and had to be redispatched (8 tasks, 1 retry, 1 fix round, review clean). The user slept through it; a repeat offender means the plan's simple bar is set too loose.
• Blocked tasks follow autopilot's rule — Still broken: under What broke and got fixed, with what was tried. Don't force past a blocker to keep the loop moving.

Override 5 — The task file is the trail

autopilot's task file (.claude/autopilot-tasks.md) is not optional here: the user is asleep and it's the only progress they can check mid-run and the only record if the session dies. Write it at the end of Phase 1 and keep it current through every state change, and append the Override 4 summary — Done, Assumptions, Deferred (unsafe), Blocked — to it when the run ends, so the file matches the report they wake up to.

Don't stall

Unattended means no waiting on human input and no infinite loops. Respect the fix-loop round cap; at the cap, stop and put remaining findings in the summary rather than looping forever. Report progress via TodoWrite when available, else reprint the checklist on state changes — counts and phase, never time estimates.