evaluation-methodology – skill Agent & Tự động hóa cho agent AI

Evaluation methodology

PluginEval scores a skill or a plugin from 0 to 100 by combining up to three layers. The static

layer is a lint. It's fast and deterministic, and it's useful for checking structure. The LLM

judge and Monte Carlo layers are experimental, not validated against human labels, so treat their

numbers as rough signals. For the trace-based eval program, see evals/README.md at the

repository root.

The judge rubric anchors are in [references/rubrics.md](references/rubrics.md). Fixes for each

anti-pattern and tips for each dimension are in

[references/improving-scores.md](references/improving-scores.md).

Evaluation depths

| Depth | Layers | Confidence label |

|---|---|---|

| quick | static | Estimated |

| standard (default for score) | static and judge | Assessed |

| deep (used by certify) | static, judge, and Monte Carlo with 50 runs | Certified |

| thorough | static, judge, and Monte Carlo with 100 runs | Certified+ |

The label names the depth that ran, not a check against human judgment. A plugin directory gets

the static layer only, whatever depth you ask for.

Static layer (lint)

The static layer reads SKILL.md and makes no model calls. It computes seven sub-scores, and the

first six feed composite dimensions:

• frontmatter_quality (feeds triggering_accuracy)
• orchestration_wiring (feeds orchestration_fitness)
• progressive_disclosure, structural_completeness, token_efficiency, ecosystem_coherence
• harness_portability

harness_portability maps to no dimension, so it doesn't change a skill's composite score. It

does carry 6% of the static layer's own score. A plugin's score is built from that layer score,

so portability findings can lower a plugin's score a little. Its findings are not counted as

anti-patterns.

The static layer flags six anti-patterns: OVER_CONSTRAINED, EMPTY_DESCRIPTION,

MISSING_TRIGGER, BLOATED_SKILL, ORPHAN_REFERENCE, and DEAD_CROSS_REF. Each flag cuts the

score by 5%, down to a floor of 50%:

penalty = max(0.5, 1.0 - 0.05 * anti_pattern_count)

The report prints a severity for each flag, but the penalty counts flags and ignores severity.

LLM judge layer (experimental)

The judge layer makes four model calls and returns four holistic scores from 0 to 1:

• triggering_accuracy: Haiku reads the description and writes 10 test prompts, 5 that should

trigger and 5 that should not. It predicts the outcome for each prompt and reports its own F1.

Nothing checks those predictions against real triggering.

• orchestration_fitness and scope_calibration: Sonnet rates the skill on a five-point rubric.
• output_quality: Sonnet imagines three tasks and rates the output it expects.

The three Sonnet calls see only the first 3,000 characters of SKILL.md. Only one judge runs,

because nothing reads the judges setting.

Monte Carlo layer (experimental)

Haiku writes 15 prompts that should trigger the skill, and the layer repeats them to reach 50

runs (100 at thorough). Each run sends the SKILL.md text and one prompt to the model, and the

layer records four measures:

• Activation rate is the share of runs with any non-empty reply, so it shows whether the model

answered, not whether the skill should have fired.

• Output quality is reply length divided by 500, capped at 1.0.
• Failure rate is the share of runs that errored.
• Token efficiency is 1 - median_tokens / 8000.

Every prompt is one that should trigger, so the layer never checks that the skill stays out of

unrelated requests. The layer's JSON includes Wilson, bootstrap, and Clopper-Pearson intervals

for its own measures. The composite ci_lower and ci_upper fields are always null.

Composite score

First, for each dimension, the engine blends the layer scores that exist, and it renormalizes

the blend weights over those layers. Second, it sums the weighted dimension scores, and it

renormalizes the dimension weights over the dimensions that have a score. Third, it multiplies

the sum by 100 and by the anti-pattern penalty.

| Dimension | Weight | Static | Judge | Monte Carlo |

|---|---|---|---|---|

| triggering_accuracy | 0.25 | 0.15 | 0.25 | 0.60 |

| orchestration_fitness | 0.20 | 0.10 | 0.70 | none |

| output_quality | 0.15 | none | 0.40 | 0.60 |

| scope_calibration | 0.12 | none | 0.55 | none |

| progressive_disclosure | 0.10 | 0.80 | none | none |

| token_efficiency | 0.06 | 0.40 | none | 0.50 |

| robustness | 0.05 | none | none | 0.80 |

| structural_completeness | 0.03 | 0.90 | none | none |

| code_template_quality | 0.02 | none | none | none |

| ecosystem_coherence | 0.02 | 0.85 | none | none |

A cell reads "none" when its layer produces no score for the dimension, even where

LAYER_BLENDS lists a weight. No layer produces code_template_quality, so it's always

unmeasured. For a plugin directory, the composite is the static layer's mean score across the

plugin's skills and agents, times 100, times the penalty for the plugin's anti-pattern count.

Badges and grades

Badges come from the composite score alone. Platinum needs at least 90, Gold at least 80, Silver

at least 70, and Bronze at least 60. Badge.from_scores accepts an Elo rating, but no command

computes one. Plugin-level badges, including the ones in the weekly CI report, come from the

static layer alone. Skill-level badges at standard depth or deeper also include the

experimental layers.

Each measured dimension gets a letter grade on the 0 to 100 scale, from A+ at 97 down to D- at

60, and F below 60.

Usage

plugin-eval score ./path/to/skill --depth quick     # static lint only
plugin-eval score ./path/to/skill                   # static and judge
plugin-eval certify ./path/to/skill                 # deep depth
plugin-eval compare ./skill-a ./skill-b             # quick depth by default
plugin-eval score ./path/to/skill --depth quick --output json --threshold 70

At standard depth or deeper, score, certify, and compare print a note on stderr that the

judge and Monte Carlo layers are experimental. For a plugin directory, the CLI prints a warning

that only the static layer runs instead. With --threshold, the command exits with code 1 when the

composite is below the value. plugin-eval init writes a corpus index, but no other command

reads it.

Examples

An abridged example of the JSON output follows. Scripts can read composite.score from it:

{
  "layers": [{"layer": "static", "score": 0.75, "sub_scores": {}, "anti_patterns": []}],
  "composite": {"score": 76.9, "ci_lower": null, "ci_upper": null, "badge": "silver",
                "confidence_label": "Estimated", "dimensions": []},
  "elo": null
}

Troubleshooting

• When a score drops after you add content, check layers[0].anti_patterns in the JSON.
• If triggering_accuracy is low at quick depth, add a trigger phrase such as "Use this skill

when" to the description, followed by several comma-separated contexts.

• Judge scores change between runs, because the model writes new test prompts and tasks each

time. Use the static layer for comparisons you want to repeat.

• If stderr says the judge could not measure some dimensions, install the LLM extra with

uv sync --extra llm.

Related

The eval-judge agent scores the four judge dimensions inside Claude Code, and the

eval-orchestrator agent runs the CLI and merges the results.