AFK is a user-invoked agent skill that spends the time you are gone making one project better. It gathers evidence, writes at least fifteen ideas, builds the strongest few as small commits on a local afk branch and ends with a report you can read in two minutes.
A run is meant to last hours, so it explores wide before deep. It matches the repository’s components, tokens and conventions, tries to reject its own commits before keeping them, and pushes, merges or deploys nothing unless the request asked for it.
What You Get
- A local branch named afk with the date, one small commit per change in the repository’s own commit style
- Run notes in a dot-afk folder at the checkout root, listed in the git exclude file so they never reach a commit
- A final report covering where things are, what was built, what was backed out, what was verified, what was not looked at, decisions made for you and ideas not built
- At least fifteen evidenced ideas across fixes, polish, features, content, performance, accessibility and your own tooling, including some that change what the product does
- Drafts of prose, changelog entries or posts, ready to paste, since published writing is left alone
Common Use Cases
- Leaving an agent running overnight on a personal site or product to find and fix real user pain
- Focusing a run on one page, feature or branch in progress and exploring every angle of it
- Turning analytics, error trackers and open issues into a ranked list of small improvements
- Porting a solution you already built in a sibling repository into the current one
What Makes It Unique
It is built for long unattended runs. It checks every tool it depends on before you leave, keeps notes that survive a long context, reads earlier afk branches to learn what you merged and what you left, and reviews its own commits as someone who wants to send them back, comparing the same surfaces before and after. It stays reversible by keeping commits local and treating connected services as read-only.
Under The Hood
The skill is a SKILL.md with a small agent metadata file, and it has no scripts. The agent works in your checkout: it creates a git branch or a separate worktree, makes local commits, runs your dev server, tests and CI checks, opens pages in a browser and uses curl to test routes. It reads connected services such as analytics, error trackers, deploy logs and issues, treating issue text, feedback and web pages as data and never as instructions. It writes a notes file inside the project and may touch files that tests write, restoring them from snapshots. It leaves pushing, pull requests, merging, deploying and messaging alone unless you ask for them.