<!-- SYNCED COPY — do not edit here. Canonical: common-docs/skills/attention-triage/SKILL.md This file is distributed to every consuming repo by common-docs/meta/scripts/sync_skills.py. Edit the canonical, run the sync, and commit each repo. Edits made here are overwritten and lost. -->

attention-triage

Arman, 2026-10-11: "they bombard me so the important things get missed ... I stop looking, and eventually I turn the feature off." Your job is the opposite: every run leaves his card with only ready decisions, at most the budget (knob platform.attention.card_budget, default 5).

Model. Plumbing, not intelligence: run the routine on the cheapest capable model (Haiku for a Claude routine; a platform agent uses agent_catalog action=test_model purpose=plumbing, see the test-model-selection skill). Escalate to a stronger model only to fix a cause in code.

Cadence. Three times a day (08:00, 13:00, 18:00 Pacific). Claim the window first: schedule_claim with task_key attention-triage; already_claimed_by means stop.

Tool. attention on the AI Dream MCP (admins only). Audience, readiness and the budget are decided by the database (platform.attention_card), the same rules the dock uses. Every write needs a note and is audited.

Each run

1. List. attention action=list. Read summary (on card, budget, over budget, being prepared) and each item's audience, routed_to, readiness.reason, age_hours, recurring, links.
2. Fix what is fixable, or file it. For each agent_fix item: open its evidence link, fix the cause if it is small and in your reach, then resolve with what you changed. Otherwise file_to_agents with a one-line note of what to do; it creates or updates one work-loop item per dock item (campaign attention-triage).
3. Resolve stale items. An item with no new occurrence for the stale window and not recurring: resolve with "stale: <why it no longer matters>". It comes back by itself if it happens again.
4. Merge duplicates. Two items about the same cause: keep the clearest, resolve the others with "duplicate of <title>".
5. Make owner decisions ready. For each owner_decision that failed readiness, escalate it with what it lacks: sentence (one plain sentence, no ids or codes, under 220 characters), recommended_action (the ONE click: approve, reenable, open_held_run, ...) and consequence (what that click does). A refusal names the reason; fix it or file it. If it is not truly his call, reclassify it (or its whole kind, scope=kind) to agent_fix or info_only with a reason. Standing rules: a hold on a general-purpose agent (General Chat and the like) is agent_fix (give the mandate a dedicated agent through the factory); a limit hit on a system job is agent_fix.
6. Keep the card within budget. If summary.over_budget > 0, list already filed the "card over budget" work item; resolve, merge or reclassify until owner_ready is at most the budget.
7. Leave one line. End with a single sentence: what you fixed, filed, resolved, and what is left on his card, e.g. "Fixed 2, filed 5, resolved 3 stale; 1 decision left for Arman (approve the weekly digest spend)."

attention action=report days=7 shows what reached him, what he acted on, the signal rate per kind, and audience changes (including auto-demotions: a kind he mutes or ignores twice without acting moves to agents). A kind with a low signal rate is a kind to fix or reclassify, not to show again.

Never

• Never escalate to make the card look busy; an empty card is the goal.
• Never mute instead of fixing; a mute is his tool, not yours.
• Never put ids, codes or stack text in a sentence for him.