Content
80%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A highly actionable, well-structured operational skill with executable commands throughout and clean one-level-deep progressive disclosure. The main gaps are missing post-action verification feedback loops for destructive/state-changing commands and some verbatim command duplication in the playbooks section.
Suggestions
Add post-action verification steps for state-changing commands (e.g., after retry-task or emit-event, re-run list-tasks/dump-task to confirm the task left the failed/sleeping state) to close the feedback loop the rubric requires for destructive database operations.
Deduplicate the "Good copy-paste playbooks" section: the "Fast path" block repeats the exact spawn/list/dump commands from the "Reproduce by spawning a task" playbook shown immediately above; consolidate them into one block with commented alternatives.
Trim the playbook section further by referencing the numbered steps in "Default debugging workflow" instead of restating the same command triples with slightly different flags.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is overwhelmingly lean, command-first material with minimal prose, but the "Good copy-paste playbooks" section repeats command blocks already shown verbatim (the "Fast path" block duplicates the preceding playbook's three commands), which could be trimmed. | 4 / 5 |
Actionability | Fully executable copy-paste commands with real flags and payloads (e.g., "spawn-task my-task -q default -P count:=42 -P enabled:=true", "emit-event order.completed -q default -P orderId=123") covering the common debug, wake, retry, and spawn cases, plus concrete rg search patterns for finding task registrations. | 5 / 5 |
Workflow Clarity | The numbered debug workflow (list-queues → list-tasks → dump-task) and per-state reasoning steps are clear, but state-changing database operations (retry-task, emit-event, cancel-task, spawn-task) have no post-action verification step to confirm the intended state change; per the rubric's cap, missing feedback loops for destructive/batch database operations caps this at 3. | 3 / 5 |
Progressive Disclosure | The single reference is clearly signaled with its scope stated ("It covers replay, decomposed checkpoints, immutable events, external idempotency, fan-out, and process-recovery tests"), points to the real file references/authoring-and-recovery.md, is one level deep with no nested references, and the operational playbook appropriately remains inline in SKILL.md. | 5 / 5 |
Total | 17 / 20 Passed |