Content
92%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 well-engineered orchestration skill: concrete commands with flags and fallbacks, a mechanism-enforced sequence with genuine error-recovery loops, and clean one-level-deep separation into real reference and script files. Its only cost is length — roughly 300 lines with a few constraints restated in multiple sections — which keeps it just short of maximal token efficiency.
Suggestions
State the work-silently rule once (in the user contract, which already declares it overrides everything) and reference it from Step 0 and Adapting instead of restating it.
Merge the never-rewrite/no-fixes rule in Guardrails with the user contract's 'No numbered fixes, no suggestions' bullet so it appears once.
The 'first time the skill runs' teaching line in Adapting duplicates the starting-line wording from the user contract; point to it rather than repeating the full sentence.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes competence — it never explains basic concepts, and its length is mostly load-bearing commands, dispatch-prompt specs, and fallbacks. However, a few constraints are repeated across sections (work silently appears in the user contract, Step 0, and Adapting; never-rewrite appears in both the user contract and Guardrails), so it is not fully lean. Above anchor 3 because the duplication is deliberate reinforcement of critical rules rather than padded explanation. | 4 / 5 |
Actionability | Guidance is fully executable: exact invocations like 'python3 scripts/feed.py serve <draft> --run <run-dir> --readers keen,skeptic --persona "keen=<one line>" ... --ready-file <run-dir>/feed.json', per-step subcommands ('feed.py next --log "needle=<-2..+2> expected... got..."', 'feed.py quit', 'feed.py transcript'), plus recall.py, signals.py, skim.py, ask.py, and room.py commands with flags and a concrete fallback (file mode) when subagents are unavailable. This is instruction-only but copy-paste actionable throughout. | 5 / 5 |
Workflow Clarity | Steps 0–7 are explicitly sequenced with hard ordering constraints enforced by mechanism ('Start the feed in the background BEFORE dispatching any reader', signals.py 'only after the timed reads are complete'), a genre-dispatch gate up front, and error-recovery loops ('feed.py close --admin ... flushes early if a reader dies', 'if the cast was wrong for the piece, recast and rerun', 'If the log cannot support an answer, the reader says so'). Checkpoints (skim-gate verdict, recall judged against intended gist) are explicit. | 5 / 5 |
Progressive Disclosure | The SKILL.md body is a workflow overview, and all detail is pushed one level deep to real, clearly signaled files: 'Read references/personas.md', 'per references/room.md', 'exactly as references/report.md specifies', 'per references/interview.md' — all four exist in the bundle, alongside a scripts/ directory whose seven files are all invoked by name. No nested references (the single interview.md → room.md mention is a sibling pointer, not a content chain). | 5 / 5 |
Total | 19 / 20 Passed |