Skill eval

Skill nikhilsitaram/claude-caliper/skills/skill-eval

Use when evaluating skill output quality, comparing skill versions, running behavioral evals, benchmarking a skill, or when triggered by /skill-evalFrom its SKILL.md

Install

npx -y skills add nikhilsitaram/claude-caliper --skill skill-eval

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

4.4 KB, as published. Nobody here has run it

Skill Evaluation

Evaluate skill output quality with assertion-based grading, blind before/after comparison, and variance analysis across 3 runs per scenario.

Setup

Resolve paths — run once at the start of any eval session:

PLUGIN_ROOT=$(python3 -c "
import json, os
p = json.load(open(os.path.expanduser('~/.claude/plugins/installed_plugins.json')))
print(p['plugins']['claude-caliper@claude-caliper'][0]['installPath'])
")
EVAL_ROOT=~/.claude/skill-evals
SKILL_EVAL_SCRIPT=$PLUGIN_ROOT/skills/skill-eval/scripts/run_eval.py
AGGREGATE_SCRIPT=$PLUGIN_ROOT/skills/skill-eval/scripts/aggregate_benchmark.py

Identify target and what "before" represents:
- New skill: before = no skill (type: "none")
- Improvement: before = snapshot of previous version
- Snapshot before editing: cp -r skills/{name}/SKILL.md $EVAL_ROOT/{name}/snapshot-before/SKILL.md
Check for evals.json at $EVAL_ROOT/{name}/evals.json. If none, create interactively — 2-3 realistic prompts including adversarial scenarios (deadline pressure, "skip this step", ambiguous requirements). Use the schema in references/schemas.md: top-level object with skill_name (string) and evals (array). Each eval needs integer id, name, prompt, and expectations.

Write config.json for this iteration at $EVAL_ROOT/{name}/iteration-N/config.json:

{
  "before": {"label": "v1", "type": "skill", "skill_path": "~/.claude/skill-evals/{name}/snapshot-before/SKILL.md"},
  "after": {"label": "v2", "type": "skill", "skill_path": "skills/{name}/SKILL.md"}
}

See: references/schemas.md for all JSON formats.

Run Evals

Spawn both variants simultaneously — don't run before first then after:

# After variant
python3 $SKILL_EVAL_SCRIPT \
  --evals-path $EVAL_ROOT/{name}/evals.json \
  --output-dir $EVAL_ROOT/{name}/iteration-N/ \
  --variant after \
  --skill-path skills/{name}/SKILL.md \
  --runs 3

# Before variant
python3 $SKILL_EVAL_SCRIPT \
  --evals-path $EVAL_ROOT/{name}/evals.json \
  --output-dir $EVAL_ROOT/{name}/iteration-N/ \
  --variant before \
  --skill-path $EVAL_ROOT/{name}/snapshot-before/SKILL.md \
  --runs 3

For no-skill baseline, omit --skill-path.

Each eval creates: $EVAL_ROOT/{name}/iteration-N/eval-{id}-{slug}/{variant}/run-{n}/output.txt and timing.json.

Draft Assertions

While runs execute, draft or refine assertions in evals.json. Good assertions are objectively verifiable, descriptively named, and test behavioral outcomes (not surface compliance). The grader flags trivially-satisfied assertions.

Grade

See: agents/grader.md

Spawn a grader subagent per eval. For each run, the grader reads output.txt, evaluates each assertion with cited evidence, self-critiques assertion quality, and writes grading.json per run directory.

PASS requires genuine task completion with cited evidence, not keyword matching.

Aggregate

python3 $AGGREGATE_SCRIPT \
  $EVAL_ROOT/{name}/iteration-N \
  --skill-name {name}

Produces benchmark.json (pass_rate, time with mean +/- stddev, delta) and benchmark.md.

Analyze Benchmark

See: agents/analyzer.md (Mode 2: Benchmark Pattern Analysis)

Spawn analyzer subagent to surface patterns aggregates hide: non-discriminating assertions (always pass), high-variance evals (flaky), time tradeoffs.

Blind Comparison

See: agents/comparator.md

Spawn comparator subagent with representative outputs from before and after WITHOUT revealing which is which. Scores Content (1-5) + Structure (1-5) -> Overall (1-10). Saves comparison.json.

Post-Hoc Analysis

See: agents/analyzer.md (Mode 1: Post-Hoc Comparison Analysis)

Spawn analyzer with comparison results + both SKILL.md files. Unblinds results, identifies why winner won, scores instruction-following (1-10), generates prioritized improvement suggestions. Saves analysis.json.

Report

Present to user:

Pass rates per variant with variance (mean +/- stddev)
Blind comparison winner + reasoning
Improvement suggestions (prioritized High/Medium/Low)
Assertion quality feedback (strengthen/drop recommendations)

Iterating

Edit skill -> snapshot as new "before" -> rerun into iteration-N+1/ -> repeat until satisfied.