evaluation-methodology（评估方法）

Evaluation Methodology

This document is the authoritative reference for how PluginEval measures plugin and skill quality. It covers the three evaluation layers, all ten scoring dimensions, the composite formula, badge thresholds, anti-pattern flags, Elo ranking, and actionable improvement tips.

Related: Full rubric anchors

The Three Evaluation Layers

PluginEval stacks three complementary layers. Each layer produces a score between 0.0 and 1.0 for each applicable dimension, and later layers override or blend with earlier ones according to per-dimension blend weights.

Layer 1 — Static Analysis

Speed: < 2 seconds. No LLM calls. Deterministic.

The static analyzer (layers/static.py) runs six sub-checks directly against the parsed SKILL.md:

These six sub-checks feed directly into six of the ten final dimensions (via STATIC_TO_DIMENSION mapping). The remaining four dimensions — output_quality, scope_calibration, robustness, and part of triggering_accuracy — receive no static contribution and rely entirely on Layer 2 and/or Layer 3.

Anti-pattern penalty is applied multiplicatively to the Layer 1 score:

penalty = max(0.5, 1.0 − 0.05 × anti_pattern_count)

Each additional detected anti-pattern reduces the score by 5%, flooring at 50%.

Layer 2 — LLM Judge

Speed: 30–90 seconds. One or more LLM calls (Sonnet by default). Non-deterministic.

The eval-judge agent reads the SKILL.md and any references/ files, then scores four dimensions using anchored rubrics (see references/rubrics.md):

1. Triggering accuracy — F1 score derived from 10 mental test prompts
2. Orchestration fitness — Worker purity assessment (0–1 rubric)
3. Output quality — Simulates 3 realistic tasks; assesses instruction quality
4. Scope calibration — Judges depth and breadth relative to the skill's category

The judge returns a structured JSON object (no markdown fences) that the eval engine merges into the composite. When judges > 1, scores are averaged and Cohen's kappa is reported as an inter-judge agreement metric.

Layer 3 — Monte Carlo Simulation

Speed: 5–20 minutes. N=50 simulated Agent SDK invocations (default). Statistical.

Monte Carlo runs N real prompts through the skill and records:

• Activation rate — Fraction of prompts that triggered the skill
• Output consistency — Coefficient of variation (CV) across quality scores
• Failure rate — Error/crash fraction with Clopper-Pearson exact CIs
• Token efficiency — Median token count, IQR, outlier count

The Layer 3 composite formula:

mc_score = 0.40 × activation_rate
         + 0.30 × (1 − min(1.0, CV))
         + 0.20 × (1 − failure_rate)
         + 0.10 × efficiency_norm

where efficiency_norm = max(0, 1 − median_tokens / 8000).

Composite Scoring Formula

The final score is a weighted blend across all three layers for each dimension, then summed:

composite = Σ(dimension_weight × blended_dimension_score) × 100 × anti_pattern_penalty

Dimension Weights

Layer Blend Weights

Each dimension draws from different layers at different ratios. With all three layers active (--depth deep or certify):

At --depth standard (static + judge only), blends are renormalized to drop the Monte Carlo column. At --depth quick (static only), all weight falls on Layer 1.

Blended Score Calculation

For a given depth, the blended score for dimension d is:

blended[d] = Σ( layer_weight[d][layer] × layer_score[d][layer] )
             ─────────────────────────────────────────────────────
             Σ( layer_weight[d][layer] for available layers )

This normalization ensures that skipping Monte Carlo at standard depth doesn't artificially deflate scores.

Interpreting Dimension Scores

Each dimension score is a float in [0.0, 1.0]. The CLI converts it to a letter grade:

When reading a report, focus first on the lowest-graded dimension that has the highest weight. A D in triggering_accuracy (weight 0.25) costs far more than a D in ecosystem_coherence (weight 0.02).

Confidence intervals appear in the report when Layer 2 or Layer 3 ran. Narrow CIs (± < 5 points) indicate stable scores. Wide CIs suggest inconsistency — often caused by an ambiguous description or instructions that work for some prompt styles but not others.

Quality Badges

Badges require both a composite score threshold AND an Elo threshold (when Elo is available). The Badge.from_scores() logic checks composite first, then Elo if provided:

The Elo threshold is skipped when Elo has not been computed (i.e., at quick or standard depth without certify). A skill can earn a badge on composite score alone in those cases.

Anti-Pattern Flags

The static analyzer detects five anti-patterns. Each carries a severity multiplier that feeds into the penalty formula.

OVER_CONSTRAINED

Trigger: More than 15 occurrences of MUST, ALWAYS, or NEVER in the SKILL.md.

Problem: Overly prescriptive instructions reduce model flexibility, increase token overhead, and signal that the author is trying to micromanage every output rather than providing principled guidance.

Fix: Audit every MUST/ALWAYS/NEVER. Replace directive language with explanatory framing where possible. Reserve hard constraints for genuine safety or correctness requirements. Target fewer than 10 such directives per 100 lines.

EMPTY_DESCRIPTION

Trigger: The frontmatter description field is fewer than 20 characters after stripping.

Problem: Without a meaningful description, the Claude Code plugin system cannot determine when to invoke the skill. The skill becomes invisible to autonomous invocation.

Fix: Write a description of at least 60–120 characters that includes:

• A "Use this skill when..." or "Use when..." trigger clause
• Two or more concrete contexts separated by commas or "or"

MISSING_TRIGGER

Trigger: The description does not contain "use when", "use this skill when", "use proactively", or "trigger when" (case-insensitive).

Problem: Even a long description is useless for autonomous invocation if it doesn't include a clear trigger signal. The system's routing model needs an explicit cue.

Fix: Prepend "Use this skill when..." to the description, followed by specific scenarios. Example: "Use this skill when measuring plugin quality, interpreting score reports, or explaining badge thresholds to a team."

BLOATED_SKILL

Trigger: SKILL.md exceeds 800 lines AND the skill has no references/ directory.

Problem: A monolithic SKILL.md forces the entire document into context on every invocation, wasting tokens on content only needed in edge cases.

Fix: Create a references/ directory and move supporting material there:

• Detailed rubrics → references/rubrics.md
• Extended examples → references/examples.md
• Configuration reference → references/config.md

The SKILL.md should link to these files with [text](references/filename.md) so the model can fetch them on demand.

ORPHAN_REFERENCE

Trigger: SKILL.md contains a markdown link [text](references/filename) where filename does not exist in the references/ directory.

Problem: Dead links waste tokens on context that will never resolve and confuse the model.

Fix: Either create the missing reference file or remove the dead link.

DEAD_CROSS_REF

Trigger: SKILL.md references another skill or agent by relative path and that path cannot be resolved from the skills/ directory.

Problem: Broken ecosystem links undermine the plugin's coherence score and may cause the model to attempt navigation to non-existent files.

Fix: Verify the referenced skill exists. Update the path or remove the reference.

Elo Ranking

PluginEval uses an Elo/Bradley-Terry rating system to rank a skill against the gold corpus.

Starting rating: 1500 (the corpus median by convention).

K-factor: 32 (standard for moderate-stakes ratings).

Expected score formula (standard Elo):

E(A vs B) = 1 / (1 + 10^((B_rating − A_rating) / 400))

Rating update after each matchup:

new_rating = old_rating + 32 × (actual_score − expected_score)

where actual_score is 1.0 for a win, 0.5 for a draw, 0.0 for a loss.

Confidence intervals are computed via 500-sample bootstrap, reported as 95% CI. Corpus percentile reflects pairwise win rate against the gold corpus. Position bias check: Pairs are evaluated in both orders; disagreements are flagged.

The plugin-eval init command builds the corpus index from a plugins directory:

plugin-eval init ./plugins --corpus-dir ~/.plugineval/corpus

CLI Reference

Score a skill (quick static analysis only)

plugin-eval score ./path/to/skill --depth quick

Returns Layer 1 results in < 2 seconds. Useful for fast feedback during authoring.

Score with LLM judge (default)

plugin-eval score ./path/to/skill

Runs static + LLM judge (standard depth). Takes 30–90 seconds.

Score with full output as JSON

plugin-eval score ./path/to/skill --output json

Emits structured JSON including composite.score, composite.dimensions, and layers[0].anti_patterns. Suitable for CI integration:

plugin-eval score ./path/to/skill --depth quick --output json --threshold 70
# exits with code 1 if score < 70

Full certification (all three layers + Elo)

plugin-eval certify ./path/to/skill

Runs static + LLM judge + Monte Carlo (50 simulations) + Elo ranking. Takes 15–20 minutes. Assigns a quality badge. Use before publishing a skill to the marketplace.

Head-to-head comparison

plugin-eval compare ./skill-a ./skill-b

Evaluates both skills at quick depth and prints a dimension-by-dimension comparison table. Useful for deciding between two implementations or measuring improvement before/after a rewrite.

Initialize corpus for Elo

plugin-eval init ./plugins

Builds the local corpus index at ~/.plugineval/corpus. Required before Elo ranking works.

Scripting the Composite Formula

Reproduce the composite score offline (pre-commit hook, CI gate):

def composite_score(dimension_scores: dict, anti_pattern_count: int = 0) -> float:
    """Replicate the PluginEval composite formula."""
    WEIGHTS = {
        "triggering_accuracy":    0.25,
        "orchestration_fitness":  0.20,
        "output_quality":         0.15,
        "scope_calibration":      0.12,
        "progressive_disclosure": 0.10,
        "token_efficiency":       0.06,
        "robustness":             0.05,
        "structural_completeness":0.03,
        "code_template_quality":  0.02,
        "ecosystem_coherence":    0.02,
    }
    raw = sum(WEIGHTS[d] * s for d, s in dimension_scores.items())
    penalty = max(0.5, 1.0 - 0.05 * anti_pattern_count)
    return round(raw * 100 * penalty, 2)

# Example: a skill with a weak triggering score
scores = {
    "triggering_accuracy":    0.65,  # D — needs description work
    "orchestration_fitness":  0.85,
    "output_quality":         0.80,
    # … fill in remaining 7 dimensions …
}
# composite_score(scores, anti_pattern_count=1) → ~76.5

JSON Output Format

Top-level shape of --output json:

{
  "composite": { "score": 76.5, "badge": "Silver", "elo": null },
  "dimensions": {
    "triggering_accuracy": { "score": 0.65, "grade": "D", "ci_low": 0.60, "ci_high": 0.70 },
    "orchestration_fitness": { "score": 0.85, "grade": "B", "ci_low": 0.80, "ci_high": 0.90 }
  },
  "layers": [
    { "name": "static", "duration_ms": 1243, "anti_patterns": ["OVER_CONSTRAINED"] },
    { "name": "judge", "duration_ms": 48200, "judges": 1, "kappa": null }
  ]
}

Parse composite.score in CI to gate deployments:

score=$(plugin-eval score ./my-skill --output json | python3 -c "import sys,json; print(json.load(sys.stdin)['composite']['score'])")
if (( $(echo "$score < 70" | bc -l) )); then
  echo "Quality gate failed: score $score < 70"
  exit 1
fi

Tips for Improving a Skill's Score

Work through dimensions in weight order. The largest gains come from fixing the top-weighted dimensions first.

Which Dimension to Improve First

Use this table when a score report shows multiple D/F grades and you need to prioritize effort.

Rule of thumb: Fix triggering_accuracy before anything else — at weight 0.25 it delivers more composite-score gain per hour than all low-weight dimensions combined.

Triggering Accuracy (weight 0.25)

• Include "Use this skill when..." followed by 3–4 comma-separated specific contexts.
• Add "proactively" if the skill should auto-activate without an explicit user request.
• Mental test: write 5 prompts that should trigger it and 5 that should not — does your description discriminate? If not, add or tighten the context phrases.

Orchestration Fitness (weight 0.20)

• Document what the skill *receives* and what it *returns* — not what it orchestrates.
• Avoid "orchestrate", "coordinate", "dispatch", "manage workflow" in SKILL.md.
• Include an "Output format" section and 2+ code blocks showing concrete worker behavior.

Output Quality (weight 0.15)

• Give specific, actionable instructions — not just goals.
• Cover at least one edge case explicitly (empty input, malformed data, etc.).
• Include an examples section showing representative inputs and expected outputs.
• The more concrete the instructions, the higher the judge will score this dimension.

Scope Calibration (weight 0.12)

• Target 200–600 lines. Below 100 is a stub; above 800 without references/ is bloat.
• Move background reading, extended examples, and reference tables to references/.
• Very narrow skills should be merged with a sibling; very broad ones should be split.

Progressive Disclosure (weight 0.10)

• Add a references/ directory (earns 0.15–0.25 bonus) and keep SKILL.md focused on the execution path. An assets/ directory adds a further bonus.

Token Efficiency (weight 0.06)

• Audit MUST/ALWAYS/NEVER count. Target < 1 per 10 lines.
• Consolidate near-duplicate bullet points and repeated-structure tables.

Robustness (weight 0.05)

• Add a "Troubleshooting" or "Edge Cases" section covering at least 3 failure modes.
• State what the skill returns when it cannot complete its task.

Structural Completeness (weight 0.03)

• Ensure at least 4 H2/H3 headings, 3 code blocks, an Examples section, and a Troubleshooting section.

Code Template Quality (weight 0.02)

• All code blocks must be syntactically valid and copy-paste ready with language tags.

Ecosystem Coherence (weight 0.02)

• Add a "## Related" section listing sibling skills or agents with relative paths.
• Avoid duplicating content that already exists in another skill — link to it instead.

Troubleshooting

"Score is much lower than expected after adding content"

The anti-pattern penalty compounds. Run with --output json and inspect layers[0].anti_patterns. If you have 5+ anti-patterns, the multiplier can reduce your score to 75% of its raw value regardless of how good the content is. Fix the flags first.

"triggering_accuracy is low despite a detailed description"

The _description_pushiness scorer looks for specific syntactic patterns, not just length. Verify your description contains the phrase "Use this skill when" or "Use when" (exact phrasing matters — it's a regex match). Also check that you have multiple use cases separated by commas or "or" to earn the specificity bonus.

"LLM judge scores vary significantly between runs"

This is expected for ambiguous skills. The judge generates 10 mental test prompts non-deterministically. Improve score stability by tightening the description and adding concrete examples. When judges > 1, averaged scores will be more stable. Use --depth deep with certify which runs Monte Carlo to get statistically-bounded scores.

"progressive_disclosure score is low even though the file is the right length"

Check whether the file is in the 200–600 line sweet spot. Files shorter than 100 lines score only 0.20 on this sub-check. Also confirm that references/ files are not empty — the scorer checks for non-empty reference files, not just the directory.

"compare shows my rewrite scores lower than the original"

Quick depth (--depth quick) only runs static analysis. If the rewrite moved content to references/ and shortened SKILL.md significantly, static scores for structural completeness may drop even though overall quality improved. Run --depth standard for a fairer comparison that includes the LLM judge's assessment of content quality.

References

• Full Rubric Anchors — all 4 judge dimensions

Related Agents

• eval-judge (../../agents/eval-judge.md) — the LLM judge that scores Layer 2 dimensions (triggering_accuracy, orchestration_fitness, output_quality, scope_calibration). Invoke directly when you need to re-run only the judge layer or inspect its reasoning.
• eval-orchestrator (../../agents/eval-orchestrator.md) — the top-level orchestrator that sequences all three layers, merges results, assigns badges, and writes the final report. Invoke when running a full certification pass or comparing two skills head-to-head.