The Autopilot Playbook

This is the design document for the engine that grows lifehacker.dev. It is a headless CMS in the literal sense: there is no admin dashboard, no WYSIWYG editor, no login screen. The “CMS” is a git repository plus a robot (Claude Code) that reads the repo, writes content into it, and opens pull requests. You — the human — are the publish button.

The loop

Every autopilot run is the same six steps:

flowchart TD
  A[1. Read brand + backlog] --> B[2. Pick highest-priority todo]
  B --> C[3. Research for real]
  C --> D[4. Draft in the right voice]
  D --> E[5. Screenshot + verify build]
  E --> F[6. Open PR, mark backlog done]
  F --> G{Human review}
  G -->|merge| H[GitHub Pages deploys]
  G -->|request changes| D

1. Read the brand. Before writing a word, the robot loads _data/brand/identity.yml, _data/brand/voice.yml, and _data/brand/glossary.yml. These define who the site is, the available voice profiles, and the words that are banned (when used sincerely).
2. Pick the work. It reads _data/backlog.yml, finds the highest-priority item with status: todo, and claims it.
3. Research for real. No hallucinated commands. Anything it tells you to run, it runs first. If a hack doesn’t work, it doesn’t get published — it becomes a Field Note about why.
4. Draft in voice. It writes using the voice profile the backlog item specifies (or the collection’s default), and lints the draft against the glossary.
5. Screenshot + verify. It builds the site locally, screenshots the result, and confirms nothing is broken before proposing the change.
6. Open a PR. It commits to a branch, opens a pull request describing what it did and why, and flips the backlog item to done. Then it stops.

The guardrails (why this isn’t terrifying)

The whole design rests on one rule: the robot proposes, the human disposes.

• No direct pushes to main. The deploy branch is human-only. The robot works on branches and opens PRs.
• No self-merge. It cannot approve or merge its own work.
• No secrets, no deploy keys. It can’t provision infrastructure or touch analytics.
• Honest attribution. Robot-written content is bylined Claude. Human-written content says so.
• Bugs go upstream. When the robot hits a theme bug, it files an issue on zer0-mistakes rather than papering over it.

If we ever loosen these, the Colophon says so, in bold, with a date.

The files that make it work

Run your own

Want to point a robot at your own zer0-mistakes site? The pattern is portable:

1. Copy the _data/brand/ files and rewrite them for your site’s identity and voice.
2. Seed _data/backlog.yml with a handful of ideas.
3. Adapt .claude/skills/grow-lifehacker/SKILL.md to your collections.
4. Keep the guardrails. Especially the no-self-merge one.

That’s the whole CMS. No dashboard required — just a repo, a robot, and a human who reads the diffs.

Status (2026-06-22): Autopilot is in assisted mode — a human runs each cycle by hand and reviews every PR. Fully scheduled, hands-off runs are designed but not switched on. When they are, you’ll read it here first.