skill-eval-preflight（飞行前技能评估）

Skill Eval

Use this skill as an authoring-side preflight for OpenClaw skills.

It is not a full runtime evaluator. It helps a skill author move from "this skill exists" to "this skill is structured well enough for first-pass evaluation and later regression work."

Good Requests

This skill is a good fit for requests like:

• "Set up eval files for this skill before I publish it."
• "Check whether this skill is ready for a first eval pass."
• "Review the description and tell me whether trigger coverage is clear enough."
• "Generate with-skill and without-skill static comparison artifacts for this skill."

Not A Good Fit

Do not rely on this skill alone for requests like:

• large-scale live runtime benchmarking
• scoring response quality across many real conversations
• tool-call correctness or factuality audits
• end-to-end production regression testing

Use a deeper evaluator after this step when you need those capabilities.

Best Fit

Use this skill when you need to:

• initialize evals/ files for a new or existing skill
• confirm a skill is ready for a first eval pass
• make positive and negative trigger coverage explicit
• catch placeholder content before sharing a skill
• write static run summaries and with-skill/without-skill comparison artifacts

Use a deeper evaluator after this step when you need live runtime experiments, tool-call quality checks, or richer output scoring.

Position In The Flow

Recommended sequence:

skill-vetter -> install/review -> skill-eval -> deeper runtime eval

• skill-vetter answers: "Is this skill safe enough to inspect or install?"
• skill-eval answers: "Is this skill structured well enough to evaluate seriously?"
• a deeper evaluator answers: "How well does the skill perform in practice?"

Workflow

1. Confirm the target folder is a skill directory with SKILL.md.
2. If the skill came from another repo or another person, do a safety review first.
3. If evals/ does not exist, initialize it with:

- evals/evals.json - evals/triggers.json - evals/README.md

1. Replace placeholder prompts with realistic authoring examples.
2. Run the readiness check before any deeper benchmarking.
3. If readiness fails, fix the missing pieces first instead of forcing a run.
4. Generate static run artifacts only after the inputs are usable.

Scripts

Initialize eval files:

python3 scripts/init_eval.py /path/to/skill

Check readiness:

python3 scripts/check_eval_readiness.py /path/to/skill

Run static eval checks:

python3 scripts/run_eval.py /path/to/skill
python3 scripts/run_eval.py /path/to/skill --check readiness
python3 scripts/run_eval.py /path/to/skill --check triggers
python3 scripts/run_eval.py /path/to/skill --check artifacts
python3 scripts/run_eval.py /path/to/skill --check files
python3 scripts/run_eval.py /path/to/skill --mode with-skill
python3 scripts/run_eval.py /path/to/skill --mode without-skill --run-group demo-baseline
python3 scripts/compare_runs.py /path/to/skill --run-group demo-baseline

Readiness Rules

A skill is ready for first-pass evaluation only when:

• SKILL.md exists
• the frontmatter description is real and not a placeholder
• evals/evals.json has at least one non-placeholder eval case
• evals/triggers.json has at least one positive and one negative non-placeholder trigger case

What This Skill Checks Well

• missing or empty eval scaffolding
• placeholder prompts that would make an eval meaningless
• missing positive/negative trigger coverage
• empty or malformed expected_artifacts
• malformed optional files declarations
• static with-skill/without-skill run artifact organization

Current Limits

run_eval.py does not perform live trigger experiments against the OpenClaw runtime. It does not score real outputs for quality, factuality, or tool correctness.

Today it performs static validation passes that:

• verify trigger files exist
• verify cases are non-placeholder
• verify positive and negative sets are both populated
• verify eval cases have usable expected_artifacts
• verify declared files entries are well-formed
• write mode-specific run summaries for later comparison

Why Publish This Skill

This skill is for authors who do not yet need a full eval lab, but do need a clean starting point. It is most useful as a lightweight preflight and scaffolding step before deeper evaluation.

Release Readiness Checklist

Before calling a skill "ready for release," aim for all of the following:

• the description names concrete trigger scenarios
• positive and negative trigger cases both exist
• placeholder content is gone
• each eval case describes observable expected artifacts
• static run summaries can be generated without errors

Compare Runs

Use compare_runs.py after both modes exist in the same run-group.

It compares:

• overall pass/fail
• per-check pass/fail
• mode-specific errors
• mode-specific notes

It writes comparison artifacts under the run-group root.

References

Read references/eval_format.md when you need the expected file formats and field meanings.