battle

Skill Battlefield

Stress-test any Claude Code skill: generate scenarios, evaluate, sharpen, and report.

Argument Parsing

Parse $ARGUMENTS:

• First positional argument = skill_path (required)
• --scenarios N = num_scenarios (default 20)
• --iterations N = max_iterations (default 5)

If no skill_path provided, print usage and stop:

Usage: /battle [--scenarios 20] [--iterations 5]

Resolve ~ in skill_path to the absolute home directory path.

Validation

1. Verify skill_path exists and is readable
2. Verify the file contains markdown content (not binary, not empty)

If invalid, print an error message describing the problem and stop.

Setup

1. Extract skill_name from the parent directory name (e.g., my-skill/SKILL.md -> my-skill)
2. Set skill_dir to the parent directory of skill_path
3. Create timestamp via date +%Y%m%d-%H%M%S
4. Create run directory: ~/.skill-battlefield/runs/{skill_name}/{timestamp}/
5. Create {run_dir}/iterations/ subdirectory
6. Copy the skill file to {run_dir}/skill-original.md
7. Print: Starting battle for {skill_name}. Run directory: {run_dir}

Set phases_dir to the phases/ directory relative to this skill file. Set refs_dir to the references/ directory relative to this skill file.

Phase 01 — Analyze

Read {phases_dir}/01-analyze.md and {refs_dir}/contracts.md.

Dispatch a subagent with:

• Full content of 01-analyze.md as instructions
• Content of contracts.md (analysis.json schema)
• run_dir and skill_dir paths

HARD GATE: Verify {run_dir}/analysis.json exists after completion. If missing, stop with error.

Print: Phase 01 complete: analysis.json written

Phase 02 — Generate Scenarios

Read {phases_dir}/02-generate.md, {refs_dir}/contracts.md, {refs_dir}/taxonomy.md, and {refs_dir}/rubric-guide.md.

Dispatch a subagent with:

• Full content of 02-generate.md as instructions
• Content of contracts.md, taxonomy.md, and rubric-guide.md
• run_dir, skill_dir, and num_scenarios

HARD GATE: Verify {run_dir}/scenarios.jsonl exists. Count lines and confirm it matches num_scenarios. If invalid, stop with error.

Print: Phase 02 complete: {num_scenarios} scenarios generated

Phase 03 — Baseline Evaluation

Read {phases_dir}/03-baseline.md, {refs_dir}/contracts.md, and {refs_dir}/rubric-guide.md.

Dispatch a subagent with:

• Full content of 03-baseline.md as instructions
• Content of contracts.md and rubric-guide.md
• run_dir, skill_dir, and skill_version = "v0"

HARD GATE: Verify {run_dir}/baseline-evals.jsonl exists. If missing, stop with error.

Calculate baseline score: mean of all score fields from baseline-evals.jsonl.

Print: Phase 03 complete: baseline score = {baseline_score}

Phase 04 — Sharpen

Read {phases_dir}/04-sharpen.md, {refs_dir}/contracts.md, {refs_dir}/taxonomy.md, and {refs_dir}/rubric-guide.md.

Dispatch a subagent with:

• Full content of 04-sharpen.md as instructions
• Content of contracts.md, taxonomy.md, and rubric-guide.md
• run_dir, skill_dir, and max_iterations

HARD GATE: Verify {run_dir}/skill-current.md exists. If missing, stop with error.

Read {run_dir}/progress.jsonl. Extract the final score from the last entry. Compute delta from baseline.

Print: Phase 04 complete: score {baseline_score} -> {final_score} (delta {delta})

Spot-Check Validation

Select 3-5 scenarios from {run_dir}/scenarios.jsonl for live validation:

• 1x scenario with the highest simulated score
• 1x scenario with the lowest simulated score
• 1-2x scenarios with scores near the threshold (score 5 or 6)
• 1x adversarial scenario (difficulty = "adversarial")

Read the final {run_dir}/skill-current.md.

For each selected scenario:

1. Run claude -p with the skill content prepended to the user_message
2. Judge the real response against the scenario's expected_behaviors using the same binary check rubric
3. Compute a real_score using the same formula as Phase 03
4. Compare simulated_score vs real_score, compute delta
5. Flag: |delta| <= 1 = OK, delta > 1 = SIMULATION-OPTIMISTIC, delta < -1 = SIMULATION-PESSIMISTIC

Write {run_dir}/spot-check.json conforming to the schema in references/contracts.md.

Print average divergence: Spot-check: avg divergence = {avg_delta}, {num_ok}/{total} checks within tolerance

Phase 05 — Report

Read {phases_dir}/05-report.md, {refs_dir}/contracts.md, and {refs_dir}/taxonomy.md.

Dispatch a subagent with:

• Full content of 05-report.md as instructions
• Content of contracts.md and taxonomy.md
• run_dir and skill_dir

HARD GATE: Verify {run_dir}/report.md and {run_dir}/skill-sharpened.md both exist. If either is missing, stop with error.

Summary

After all phases complete, read {run_dir}/report.md and extract the top findings.

Print to user:

=== Battle Complete ===
Skill: {skill_name}
Score: {baseline_score} -> {final_score} ({delta})
Top findings:
  1. {finding_1}
  2. {finding_2}
  ...
Report: {run_dir}/report.md
Sharpened skill: {run_dir}/skill-sharpened.md
Diff: diff {run_dir}/skill-original.md {run_dir}/skill-sharpened.md