AFK Mode
A modifier on the current task, not a session state. "afk, research X" means: research X deeply, autonomously, end to end. Don't reply "AFK on, ready when you are." Don't wait. Don't ask. Acknowledge in one line max, then plan and execute in the same turn.
A bare "afk" with nothing in flight → ask what to do. Otherwise apply it to whatever is in flight.
Procedure
- No preamble. One short line is fine ("Going deep, plan first."), then step 2 in the same turn.
- Plan with a todo list. Concrete, ordered, verifiable steps.
- Execute. Mark in_progress → completed.
- Self-review the diff: stray logs, half-applied refactors, unused imports, broken tests, missed edges. Fix them.
- Final report:
## AFK report
**Task:** <one-line restatement>
**Status:** Done / Done with caveats / Blocked
**What changed:** <file: one-line why>
**Verified by:** <test/manual check + result>
**Assumptions made:** <decisions taken without asking>
**Open questions:** <only if blocking>
**Suggested next:** <if applicable>
A pure-research task → swap "What changed" for "What I found", "Verified by" for "Sources".
Rules
- Ambiguity → reasonable judgment, note the assumption, continue. Surface it in the final report.
- Think hard on every non-trivial decision.
- Use subagents (a read-only explorer for codebase research, a planner for architecture, parallel agents for independent investigations).
- Verify your own work: tests, types, lint, build, a manual click-through if it's UI. No "code looks right". Can't verify → say so.
- Stay in scope. AFK = "do it thoroughly", not "rewrite the codebase".
Safety (still applies)
- No destructive git without prior authorization (force-push, reset --hard, branch -D, rm -rf on tracked paths).
- No push, PRs, or messages on the user's behalf unless the task says so.
- No committing secrets. No
--no-verify.
- Hit something irreversible and unauthorized → STOP, note it in the report.
When NOT to use
- The path forward genuinely needs user input → one question beats an hour of spinning.
- Irreversible external actions (prod deploys, paid APIs at scale) without prior authorization.
1---2name: afk-mode3description: Take a task end to end, autonomously, with no clarifying questions.4---56# AFK Mode78A modifier on the current task, not a session state. "afk, research X" means: research X deeply, autonomously, end to end. Don't reply "AFK on, ready when you are." Don't wait. Don't ask. Acknowledge in one line max, then plan and execute in the same turn.910A bare "afk" with nothing in flight → ask what to do. Otherwise apply it to whatever is in flight.1112## Procedure13141. No preamble. One short line is fine ("Going deep, plan first."), then step 2 in the same turn.152. Plan with a todo list. Concrete, ordered, verifiable steps.163. Execute. Mark in_progress → completed.174. Self-review the diff: stray logs, half-applied refactors, unused imports, broken tests, missed edges. Fix them.185. Final report:1920```21## AFK report22**Task:** <one-line restatement>23**Status:** Done / Done with caveats / Blocked24**What changed:** <file: one-line why>25**Verified by:** <test/manual check + result>26**Assumptions made:** <decisions taken without asking>27**Open questions:** <only if blocking>28**Suggested next:** <if applicable>29```3031A pure-research task → swap "What changed" for "What I found", "Verified by" for "Sources".3233## Rules3435- Ambiguity → reasonable judgment, note the assumption, continue. Surface it in the final report.36- Think hard on every non-trivial decision.37- Use subagents (a read-only explorer for codebase research, a planner for architecture, parallel agents for independent investigations).38- Verify your own work: tests, types, lint, build, a manual click-through if it's UI. No "code looks right". Can't verify → say so.39- Stay in scope. AFK = "do it thoroughly", not "rewrite the codebase".4041## Safety (still applies)4243- No destructive git without prior authorization (force-push, reset --hard, branch -D, rm -rf on tracked paths).44- No push, PRs, or messages on the user's behalf unless the task says so.45- No committing secrets. No `--no-verify`.46- Hit something irreversible and unauthorized → STOP, note it in the report.4748## When NOT to use4950- The path forward genuinely needs user input → one question beats an hour of spinning.51- Irreversible external actions (prod deploys, paid APIs at scale) without prior authorization.