Content
85%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-structured instruction skill: clear step sequence, genuine progressive disclosure to real one-level-deep reference files, robust degradation paths, and concrete interface contracts. Weaknesses are minor: some repetition around the optional image-skill dependency, and a mandated script (`gacha.py`) that is not present in the bundle with no fallback for its absence.
Suggestions
Deduplicate the optional image-skill dependency explanation: state the availability-check and fallback logic once (e.g., in '可选依赖') and have Step 5 reference it in a single line instead of restating it in Step 5, the fallback block, and '兼容性'.
Add a degradation path for a missing `gacha.py` (e.g., 'script file absent → fall back to the 10-category preset table') or ship the script in the bundle, since the body currently mandates executing it.
Compress the per-step dialogue-tone sample quotes into a short table of tone principles per step, keeping the '不要照抄' instruction, to reduce token cost without losing the persona spec.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely lean — the 10-direction table and command snippets earn their tokens — but the image-skill caveat is restated three times ("可选依赖", Step 5, "兼容性"), and the per-step dialogue-tone scripts plus the closing README note add padding that could be trimmed into one section. | 4 / 5 |
Actionability | Guidance is concrete throughout: exact invocation `python3 ${SKILL_DIR}/gacha.py [次数]`, an explicit image-skill interface contract (params, exit code 0 success, non-zero fallback), file paths, and a structured error-message template. The main gap is that `gacha.py` is declared a built-in tool and marked "必须执行脚本" yet is absent from the bundle, and the degradation table only covers 'Python 不可用', not a missing script. | 4 / 5 |
Workflow Clarity | Steps 1–6 are clearly sequenced, a trigger-judgment table disambiguates the three entry modes, every step ends in an explicit user checkpoint ("展示后…引导用户"), and the error-handling section defines feedback loops (retry once, then degrade; per-fault degradation table) plus a uniform degraded-output format. | 5 / 5 |
Progressive Disclosure | The body is a true overview: each detailed step points to a clearly signaled, one-level-deep `references/*.md` file, and all six referenced files (identity-tension, boundary-rules, naming-system, avatar-style, output-template, error-handling) exist with no nested references — the reference files only point back to SKILL.md sections. | 5 / 5 |
Total | 18 / 20 Passed |