How to use the Watchdog Continuity Checks agent skill

What the watchdog skill makes the agent do

Watchdog covers continuity checks: timed or evidence-triggered interventions that keep an agent workflow from going idle, drifting, or adding unnecessary bureaucracy. The skill defines a three-level stack. A wake is a small liveness nudge that restarts motion and does not reframe the work. A refocus is a medium alignment nudge that re-centers the agent on its role, the current approved workstream and its stop conditions without interrupting valid work. An alignment checkpoint is a larger, deliberate pause in which the agent runs a full refocus before proceeding.

The skill stresses that scheduling tools such as cron, timers and rig watchdog should not decide the action. The intervention level is chosen by evidence, not by how often a tick fires.

• Wake: owner appears idle, stale, blocked, or missing a next handoff.
• Refocus: output shows mode drift, approval regression, or weak stop-condition reasoning.
• Alignment checkpoint: phase boundary, lifecycle change, product-intent decision, or a confusing contradiction.

Install the watchdog skill

Download the watchdog ZIP from this page and unzip it into one of two places. For a personal setup, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Project only: unzip into .claude/skills/ in the project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need a special command. The skill is meant to load when your request matches its description. Ask in plain language about the thing you are doing: setting a watchdog policy, deciding how hard to nudge a stalled agent, or writing the text of a wake or refocus message.

Name the workflow and the symptom when you can. The more specific the situation, the easier it is for the agent to pick the right level.

• "Set up a watchdog policy for our long-running rig that has artifact-pool edges."
• "The producer seat has been idle after its handoff. Should I wake it or refocus it?"
• "Draft a refocus message for the reviewer that says finish the current action first."
• "We're at a phase boundary. Do we need an alignment checkpoint?"

Write good wake and refocus messages

The skill gives a checklist for refocus text. It should start with finish current action first, name the intervention kind, and name the current approved workstream or say that none is known. It should state what continuity means (continue, verify, hand off, or explicitly park) and list the stop conditions, such as scope or risk changes, a failed gate, contradictory evidence, or no continuity chain naming a next step.

A good message must not turn itself into a new approval gate, and it should not wake delivery seats unless an approved workflow exists and the owner is stale.

• Start with: finish current action first.
• Name the intervention: wake, refocus, or alignment checkpoint.
• Name the approved workstream, or say none is known.
• State continuity: continue, verify, hand off, or park.
• List stop conditions.
• Do not ask for re-approval of approved work.

Choose cadence without flooding agents

The skill separates scan cadence from wake cadence. The scan checks state often; the wake fires much less often and only while work remains actionable. The skill's worked example is a scan every 30 seconds with wakes at most every 600 seconds, and it says the numbers are illustrative. The durable rule is the separation itself.

The skill lists seven failure modes to avoid. The most common are cadence that is too frequent, waking the wrong seat, static repeated reminders that teach agents to answer the reminder instead of advancing the artifact, and refocus messages misread as a new top-priority task.

• Prefer one workflow watchdog plus targeted exceptions over many per-seat nag loops.
• Use queue nudges instead of a watchdog policy when a clear next handoff is already queued.
• Fix bad startup or workflow design rather than compensating with a watchdog.

Use artifact-pool loop edges

The skill describes one evidence-aware pattern. A consumer-pool wake fires when a ready artifact exists, waking the consumer loop head. A producer-edge repair fires when upstream completion exists but the downstream artifact is missing, waking the producer loop head.

The skill frames this as a guardrail against cold pools and missing edge artifacts, not as a general workflow engine.