Skip to content
Back to skills

Autopilot

ASecurity

Kaizen autonomous mode — takes a request all the way without stopping, through the right skills (plan or debug, work, simplification, review with fixes, capture, commit, push, PR, CI watching). A code change ends as an open PR. Use only when the user explicitly asks for end-to-end autonomous work or invokes /kaizen:autopilot — ideally after /kaizen:brainstorm. For step-by-step follow-up, use plan, work, debug.

  • 2 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added October 19, 2026
ai-agentsgobashnodegitdocumentation

Works with

  • mcp

Security analysis

A100/100

Scanned October 19, 2026

npx -y skills add Lingelo/dojo --skill autopilot --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Autopilot?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Autopilot
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/lingelo-autopilot/badge)](https://www.skillsdirectory.com/skills/lingelo-autopilot)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: autopilot
description: Kaizen autonomous mode — takes a request all the way without stopping, through the right skills (plan or debug, work, simplification, review with fixes, capture, commit, push, PR, CI watching). A code change ends as an open PR. Use only when the user explicitly asks for end-to-end autonomous work or invokes /kaizen:autopilot — ideally after /kaizen:brainstorm. For step-by-step follow-up, use plan, work, debug.
allowed-tools: Bash, Read, Write, Edit, Glob, Grep, Agent, AskUserQuestion, TaskCreate, TaskUpdate, TaskList
argument-hint: "[feature, bug, ticket or plan path]"
---

# Autopilot — from request to PR, without stopping

**Outcome:** the request reaches the end state its shape calls for, produced by the Kaizen skill whose
job it is, with everything left unresolved recorded where the user will see it. A code change ends as
an **open PR** whose URL you give, CI settled, after being implemented, simplified, reviewed (eligible
fixes applied, the rest recorded), captured if relevant, committed and pushed. **Merging stays with
the user** unless explicitly authorized.

Read `${CLAUDE_PLUGIN_ROOT}/references/conventions.md`.
`K="${CLAUDE_PLUGIN_ROOT}/scripts/kaizen.mjs"`

## Interaction

The user is **only** questioned through `/kaizen:brainstorm`, and only if they are present. Everything
else moves on without waiting: what is reversible is done and shown (the user will correct it
afterwards); only an **irreversible** action outside what was granted (merge, force push, data
deletion, deployment) stops the run. Report to the user in their language.

## Visibility

Create one task per step (`TaskCreate`) and keep them up to date. A step is only done after it really
ran; a child skill's return chains the next step **in the same turn**; the turn does not end before
DONE or a motivated stop.

## Routing

Match the request to the skill whose job it is:
- a **plan path**, or a plan written in this session → plan route (step 2 directly);
- a **concrete bug** (symptom, red test, bug ticket) → `kaizen:debug mode:return`;
- an **ambiguous product shape** (several plausible readings) → `kaizen:brainstorm mode:return` if a
  human is present, otherwise `kaizen:plan mode:return` which will record its assumptions;
- an outcome **that is not code** (ideas, explanation) → the skill concerned, and that is all;
- any other code change → `kaizen:plan mode:return`.

**No "trivial change" shortcut.** Even for five lines, the run goes through a work source (a plan,
short if need be), `work` under the gate, `node "$K" verify` and `kaizen:review`. Only simplification
(step 3) is skipped for a small diff, and delivery (9-10) without a remote. Whoever wants a change
without ceremony uses `/kaizen:work` directly, not `autopilot`. **Only exception: `lean` profile**
(`node "$K" config` → `profile`) — a change of ≤ ~30 lines with no risk surface (see `conventions.md`)
goes to `kaizen:work mode:return` with its 2 to 6 announced units, without a written plan; gate,
`verify`, review and delivery stay identical.

When in doubt, the route requiring the most evidence. **Never** an improvised plan on top of an
existing plan, nor a plan picked at random from the plans folder.

## The run (code-changing routes)

1. **Work source** — an implementation-ready plan (check `<!-- kaizen:units -->`), or a `fixed` return
   from `kaizen:debug`. `blocked`, `needs-human`, `settled-decision-invalidated` → **stop**.
2. **Implement** — `kaizen:work mode:return <plan>` (debug route: already done, go to 3). Only
   `status: complete` moves on. The quality gate stays active during the whole run
   (`node "$K" gate on --plan <plan>`, removed at the end).
3. **Simplify** — `simplify` skill if available, otherwise one pass yourself (reuse, clarity,
   efficiency); skipped for a documentation diff or one under ~10 lines. Verify again.
4. **Review** — `kaizen:review mode:agent plan:<plan>` (without `plan:` on the debug route). A finding
   showing that a settled decision **cannot work** (infeasible, wrong target, destructive) stops the
   run before any push.
5. **Apply the fixes** — confirmed P0/P1 and `gated_auto` P2: apply, verify again (`node "$K" verify`),
   commit (`fix(<JIRA>): review fixes`). Nothing stays only in the working tree. Record the review
   again with the post-fix verdict (`node "$K" review record --verdict …`): the push hook requires it.
6. **Record the rest** — every actionable finding not applied, every decision flagged along the way:
   in the PR description ("Open points" section), or in the final report if there is no PR.
7. **Capture** — `kaizen:learn mode:auto` if the run produced durable reasoning the code, tests and
   plan do not carry. "Learning not written" is a success. The learning goes into the PR.
8. **Browser tests** — UI change and a browser tool answering (the Playwright MCP server Kaizen ships, or another): walk through
   the acceptance examples touched; failure → fix, verify again.

**Delivery precondition** (from 9 on): empty `git remote` → everything stays in local commits, push, PR
and CI are skipped. That is not an error.

9. **Ship** — `kaizen:ship <plan> mode:auto`: checks, size (`node "$K" size`; above the limit, the PR
   justifies it), branch push (never the default branch, never `--force`), PR with description,
   reviewer guide, rollout and rollback, open points. An existing PR for the branch is updated, not
   duplicated.
10. **Drive the PR** — `kaizen:watch-pr <url> mode:pipeline`: review feedback handled, CI repaired (at
    most 2 fixes per cause, no test disabled, no empty commit), branch updated only if GitHub asks for
    it. It returns `looks-ready`, a motivated blocker or its residuals: record them in the PR ("Open
    points") and finish.
11. `node "$K" gate off`, then the final report and `DONE`.

**No deployment.** The run stops at the ready PR; releasing to production goes through
`/kaizen:deploy`, with the approval typed by the user for a protected environment.

## Stops (say why)

Work source impossible to produce · child return other than complete and substantiated · settled
decision invalidated · the project's delivery process not satisfied. A stop pushes nothing that was not
already pushed; it removes the gate (`gate off`) and summarizes the exact state and the possible
resume.

## Final report

```
DONE — <title>
PR: <url> — ✅ looks ready | 🟡 reservation: … | ⛔ blocked: … (watch-pr)
Plan: <path> · Units: 4/4 · Review: 1 P1 fixed, 2 P3 recorded · Learning: <path | none>
Open points: …
```

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.