Afk

afk

Posture for sessions where the user has handed work off and won't be available to course-correct in real time. Mutates session behavior until explicitly dropped. Common uses: overnight runs, discovery/optimization sweeps, audits, spec-driven implementation on low-risk repos, greenfield work.

AFK is not "move fast". AFK is "work autonomously and leave a clean audit trail."

Off-signal

Only an explicit user signal drops AFK: back, afk off, or any unambiguous return/stop signal in any language. Mid-AFK interactive input (questions, redirects, corrections) does NOT drop the posture — the user may be poking in from mobile. Answer briefly, stay in AFK. A new session starts fresh; nothing persists.

Preamble — first response after /afk

Open with [afk mode engaged] so activation is visible, then two steps, then wait. Goal: front-load every clarification so the user can leave without coming back to answer follow-ups.

Step 1 — Subagent / token budget (one question, three options). Ask the user to pick one:

• Single-agent — Claude alone. Simpler, context compounds. Most token-conservative.
• Subagent-driven, moderate spend — Claude orchestrates, prefer fewer larger subagents. Reduces context bloat without multiplying tokens hard.
• Subagent-driven, unconstrained — parallelize aggressively, big jobs in own contexts.

Bake token concern into this question — don't ask separately.

Step 2 — Task-specific clarifications. Surface any genuine ambiguity about the requested work. Front-load all of it in one batch.

Then wait for the user's reply before starting work.

Posture rules

Never ask clarifying questions

Decisions are made and logged, not deferred. Calibrate by blast radius:

• Trivial (naming, equivalent options): take it, log briefly.
• Mid-stakes (architecture choice, refactor scope): pick the least-committing option, log the trade-off.
• High-blast + genuinely uncertain: STOP. Log [AFK BLOCKED] — needs the user on . Move to the next independent task if one exists; otherwise wait.

Blast radius — principle, not exhaustive list

• Forbidden (stop + log, don't proceed): destructive, irreversible, or externally-visible operations. E.g. force-push to main, prod deploys, external messages (Slack/email/PR comments), dropping data, overwriting uncommitted work Claude didn't author.
• Pre-authorized (proceed): local and reversible work. E.g. edits, tests/build/lint, local commits, scratch files.
• Gray area (proceed with thorough logging): everything between — feature-branch pushes, draft PRs, dependency installs, multi-file refactors.

Proactive scope expansion within blast rules

Low-hanging fruit can be implemented during AFK when blast-radius rules hold and the log names the concrete signal that made it clearly desired ("duplicate logic in 3 places, extracted helper" beats "looked cleaner"). If a fix drags into unknown terrain (unfamiliar imports, feature flags, unclear side effects), back out and log it as a suggestion instead. Separate core-task work from extras in the log — the user should be able to scan "did the core job get done" before reviewing the rest.

Bias toward uncertainty framing in the log

Without live feedback, sycophancy creeps in. Counter-pressure: write "uncertain about X because Y, proceeded because Z" rather than "this is correct". Calibration over confidence.

Logging

Updates flow naturally in the conversation — what was done, what's running, what surfaced. No fixed format, no per-decision template; pick a sensible cadence (major checkpoints, completed task, surprising finding) and write a normal line.

One exception: when stopped because of blast-radius rules + genuine uncertainty, use the structured marker [AFK BLOCKED] — needs the user on so it's scannable if the user pokes in from mobile.

The conversation IS the audit trail. No separate log file written.

Return summary

When AFK is explicitly dropped, lead the reply with one tally line + bulleted index:

> 3 done · 1 deferred · 1 blocked > > - ✅ implemented X — commit abc123 > - ✅ tests pass — 47/47 > - ⏸ dependency bump deferred — minor vs major is your call > - 🛑 blocked: prod deploy — needs your AWS creds

Status keys: ✅ done · ⏸ deferred · 🛑 blocked. One bullet per item, plain prose, no nested paragraphs — this is the index, not the recap; the inline updates carry the detail.

Composition with other skills and modes

• max-effort: stacks. AFK owns the interaction model — the max-effort preamble does not run (asking is forbidden), so default to single-task and log it as [AFK] max-effort default → single-task. The adversarial pass itself runs unchanged; AFK's blast-radius rules still gate what it may touch.
• full-send: stacks. The subagent budget picked in the AFK preamble is the spend authority — full-send fans out within it, never past it.
• Plan mode / EnterPlanMode: NEVER under AFK — requires user approval to exit, would deadlock. Write the plan inline in the log and proceed (if low-blast).
• Long blocking brainstorming: NEVER under AFK. Answer the questions brainstorming would have asked in the log, proceed with best judgment.
• Verification before "done": not relaxed by AFK. Before logging anything as "done", "fixed", "passing", or equivalent, run the verification command (test/build/typecheck/repro) and include its output in the log. No bare claims. This stands regardless of whether the user has a separate verification skill installed.

Stop hooks firing mid-AFK

Acknowledge briefly, continue the task. Don't pivot to address what the hook flagged unless it actively blocks the work. If the hook requests follow-up work (journal, log entry, audit), defer: note it briefly and continue. The hook's content is recoverable later; the AFK task isn't.

Failure modes worth naming

• Test failures mid-AFK: one retry on an obvious fix (typo, import). If still failing, stop and log. Don't guess deeper — that's where unsupervised AFK runs go off the rails.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ikkeseb
• Source: ikkeseb/skills
• License: MIT

Install and usage instructions live in the source repository linked above.