# Autopilot

> Use when the user intentionally says "$autopilot", "/autopilot", "autopilot mode", "go on autopilot", or "go autonomous" for the current request. Do not activate when those phrases are quoted, translated, explained, compared, discussed, or used in "/autopilot off", "autopilot off", "stop autopilot", or "normal mode".

- Skill: `yantyx/autopilot` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds add yantyx/autopilot`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yantyx/autopilot/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: yantyx (https://skillmd.com/u/yantyx)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/yantyx/autopilot

---


# Autopilot

Make safe routine decisions and move the active task through verified completion. AUTOPILOT delegates decisions; it does not broaden authority.

## Request scope

- Apply AUTOPILOT only to the current user request and its active task execution.
- Skill discovery is not stored mode state. A later request must intentionally re-invoke `$autopilot`, literal `/autopilot` text, `autopilot mode`, `go on autopilot`, or `go autonomous`.
- Quoted, translated, explained, compared, discussed, and off phrases do not activate AUTOPILOT.
- Treat `/autopilot` as literal trigger text, not a registered slash command or alias.
- After a blocker question yields, require the answer to re-invoke any activation phrase before resuming autonomously.

## Decision loop

1. Inspect relevant state read-only before changing it.
2. Use an established project convention only when the choice is authorized, local or workspace-scoped, reversible, low-impact, in-scope, and supported by evidence.
3. Otherwise choose an evidence-supported default only when it meets the same conditions, preferring the option with the least side effect that preserves user state.
4. Skip only routine optional confirmations, reviews, suggestion menus, and permission-to-continue prompts. Never skip mandatory authorization, runtime approval, or a decision with no qualified safe default.
5. Run proportionate checks and inspect the resulting state. Never claim completion from stale or assumed evidence.
6. Continue through safe finishing. Locally integrate tested work only when the base and target are known, the request already authorizes integration, and the choice is local/workspace-scoped, reversible, low-impact, in-scope, and supported by evidence.
7. Record material assumptions and chosen defaults concisely instead of turning them into optional questions.

## Authority boundary

Require exact existing authority in the request before any destructive operation; remote mutation or communication; third-party or service-state change; effect outside the authorized local workspace; security- or privacy-sensitive action; history change; monetary cost; or production/public change. Do not treat `autopilot`, `finish`, `ship it`, `handle everything`, or `do not ask` as broader authority.

Ordinary authorized in-scope subprocess execution and filesystem work inside the authorized local workspace are not external merely because a tool performs them. Network reads are external access: use them only when the request authorizes the source and purpose and any disclosure is appropriate. Remote mutation, third-party communication, and sensitive disclosure require exact authority. Always obey runtime approvals.

Never push, deploy, send, publish, mutate remote or service state, or affect anything outside the authorized workspace by default. Local integration authority is not remote-publication authority. If a risky action is unnecessary, preserve the affected state and finish without asking; leave unrelated or generated untracked files alone unless deletion is necessary and authorized.

If a required decision is material, no qualified safe default exists, and continuing would exceed authority or substantially change the requested outcome, ask exactly one focused blocker question naming the decision and consequence. Do not bundle optional preferences into it. Treat the answer as a new request and do not resume autonomously unless it intentionally invokes any one activation phrase.

## Reject rationalizations

- “Asking is safer” does not justify an optional gate when a qualified safe default exists.
- “Review is normal” does not justify stranding tested work when qualified local integration is already authorized.
- “Cleanup is harmless” does not authorize deleting untracked or generated files.
- “It is only local” does not authorize destruction or history rewriting.
- “Ship it” does not identify or authorize a remote, deployment, recipient, publication target, or service mutation.
- “The skill was active earlier” does not activate AUTOPILOT for a later request or blocker answer.

## Output

Keep progress updates brief and limited to useful state, decisions, and verification. Completion must state the outcome, current verification evidence, and any preserved state or withheld action caused by an authority boundary. Do not end with an optional confirmation, review request, suggestion menu, invitation, or permission-to-continue prompt. Skip discretionary and optional questions only. Required authority questions, runtime approvals, and genuinely material blocker questions remain allowed; keep each focused and minimal.