Content
81%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.
The body is highly actionable and workflow-safe: executable examples for every tool, explicit feedback loops for background waits, collision avoidance for parallel runs, and a pre-dispatch checklist. Its weaknesses are repetition of the wait/kill guidance across several sections and a monolithic single-file structure that could offload advanced material to reference files.
Suggestions
State the kill/collect rule once (e.g., in the wait_for_subagents section or the Checklist) and reference it elsewhere instead of repeating "use kill=True only when the delegated work is no longer needed" and "collect final results before finishing" in the timeout table, Handling `running` status, result-format bullets, and both example comments.
Merge "Handling `running` status" into the timeout/kill subsections it restates, keeping only the new information (doing independent work while waiting, sub-agents continuing during user questions).
Move the advanced checkpoint-pattern section, the prompt-brief template, and the per-parameter API reference into one-level-deep reference files (e.g. references/prompt-template.md, references/checkpoint-patterns.md) so SKILL.md stays a lean overview, and add clearly signaled links to them.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Almost all content is tool-specific knowledge Claude cannot already know (session_id semantics, fork/resume lifecycle, POSIX timeout behavior, pattern matching), so there is no padding with known concepts — but the skill is noticeably repetitious: the rule "use kill=True only when the delegated work is no longer needed" plus "collect and process final results before finishing the parent task" appears at least five times (timeout table, Handling `running` status, result-format bullets, Parallel Example comments, Checkpoint rules), and "Handling `running` status" restates the timeout table directly above it. This is more than the "minor instances" of anchor 4, fitting anchor 3's "could be tightened". | 3 / 5 |
Actionability | Fully executable, copy-paste-ready code covers every tool and the common cases: `find_agents` keyword search and full-list browsing, `call_subagent` sync/parallel/fork/checkpoint variants, a reusable `dispatch()` helper, and `wait_for_subagents` usage, plus concrete Good/Bad prompt examples with an explicit parallel-context template. This clearly matches the anchor for fully executable guidance with examples covering common cases, not anchor 4's "minor gaps". | 5 / 5 |
Workflow Clarity | The multi-step process is clearly sequenced (when to delegate → select target → dispatch → wait → handle `running` → report) with explicit validation and feedback loops: timeout is defined as "still running" not failure, the parent must read `Last message:` and keep waiting or explicitly kill, parallel write collisions are guarded ("Never let multiple sub-agents write to the same file concurrently"), and a pre-dispatch Checklist verifies each step. This matches the anchor-5 example (explicit validation, error-recovery loops, checklists) rather than anchor 4's minor validation gaps. | 5 / 5 |
Progressive Disclosure | The single file is well-organized with clear section headers, per-parameter reference subsections, worked examples, and a closing checklist, making navigation easy — but at ~430 lines with no bundle files, advanced material (checkpoint pattern matching, the prompt-brief template, the full parameter API reference) could be split into one-level-deep reference files to keep SKILL.md a leaner overview. Good structure with minor organization gaps fits anchor 4; not 3 because nothing is buried or clearly misplaced, not 5 because it is a monolithic file rather than an overview with well-signaled references. | 4 / 5 |
Total | 17 / 20 Passed |