AFK

afk — a simple coding flow for Claude Code

One help router, one optional end-to-end orchestrator, focused workflow skills, a packaged implementation agent pair, and a persistent brain/ memory vault the flow reads before acting and writes back to as it learns. Just the steps that matter.

Full documentation: https://alexanderop.github.io/afk/

Install

Inside any Claude Code session:

/plugin marketplace add alexanderop/afk
/plugin install afk@afk

Verify with /help — the skills appear under the afk: namespace. Update later with /plugin marketplace update afk.

To try it without installing:

git clone https://github.com/alexanderop/afk
claude --plugin-dir ./afk

/afk:qa needs Vercel's agent-browser CLI installed and available on your PATH when it routes to frontend browser QA.

Skills

The four-step coding flow: grill → implement → simplify → qa. Use /afk:ship to drive the planning-through-verdict loop, or grab any skill standalone. /afk:qa is the closer for the implement path: it verifies the change with real evidence and produces a ship/no-ship verdict (/afk:ship stops at that verdict). /afk:batch is the fan-out alternative to implement: one PR per independent unit, run in parallel worktrees. /afk:prototype builds a throwaway exploration before committed planning.

The brain/ vault skills (init-brain, brain, map-codebase, reflect, ruminate, meditate, plan, review) keep a persistent Obsidian-compatible store of your project's principles, gotchas, and decisions — wired into the flow automatically. /afk:init-brain also authors a tight CLAUDE.md (with an AGENTS.md symlink) that onboards any agent to the repo and points at the brain.

For per-skill detail see the Reference section of the docs.

If the skills don't auto-trigger

Invoke directly: /afk:help, /afk:ship, /afk:grill, /afk:implement, /afk:batch, /afk:simplify, /afk:qa, /afk:prototype, /afk:write-good-goal, /afk:research, and the brain skills /afk:brain, /afk:init-brain, /afk:map-codebase, /afk:reflect, /afk:ruminate, /afk:meditate, /afk:plan, /afk:review. Run /doctor to check plugin loading.

What lands in your repo

qa/                          # QA reports + screenshots (gitignored)
brain/                       # persistent memory vault — everything AFK writes lives here:
brain/context.md             #   domain glossary, grown by /afk:grill
brain/decisions/NNNN-*.md    #   ADRs worth recording, offered sparingly by grill
brain/plans/<slug>.md        #   the agreed plan, input to /afk:implement
brain/prototypes/            #   prototype verdict notes for throwaway exploration
brain/principles/            #   principles + gotchas the flow reads before acting
brain/codebase/              #   as-is maps of how areas work, by /afk:map-codebase

Testing the plugin

See docs/testing-strategy.md for the full approach.

bun run test              # zero-token unit + integration checks (run on every edit)
bun run test:e2e          # one cheap headless turn (~$0.01): plugin actually loads
bun run test:evals        # model-backed behavioral evals via claude -p
bun run test:triggers     # model-backed: do skills fire from bare prompts? (pre-release only)

Credits

• The grilling interview prompt is based on Matt Pocock's "interview me about the plan" technique, extended with domain-glossary and ADR awareness.
• Skill structure and red-flags tables borrowed from obra/superpowers.
• The brain/ memory vault — its skills (brain, init-brain, reflect, ruminate, meditate, plan, review) and hooks — is derived from brainmaxxing by Lauren Tan (MIT), rewritten to AFK's skill conventions.