swarm: Skill for Claude Code

EXECUTE NOW

Target: $ARGUMENTS

Parse immediately:

• If target is a run_id: synthesize that specific run
• If target is empty: list recent runs without experiment notes and ask which to synthesize
• If target is --all: synthesize all runs that lack experiment notes
• If target is --batch N: synthesize the N most recent unsynthesized runs

Execute steps 0 through 6 in order. Do not ask for confirmation between steps.

Step 0: Validate Target

1. Check that runs/{run_id}/ exists
2. Check that runs/{run_id}/run.yaml exists — if not, run python scripts/backfill-run-yaml.py --run-id {run_id} to generate it
3. Check if vault/experiments/{run_id}.md already exists — if so, report "Already synthesized" and stop (unless --force)

Step 1: Read Run Metadata

Read runs/{run_id}/run.yaml and extract:

• slug, created_utc
• experiment.type, experiment.hypothesis, experiment.swept_parameters, experiment.seeds, experiment.total_runs
• results.status, results.primary_metric, results.primary_result, results.significant_findings
• artifacts.* — what files are available
• tags
• links.claims — any pre-linked claims

If experiment.type is sweep or study, also read the summary JSON file referenced in artifacts.summary.

Step 2: Read Summary Data

Based on experiment.type:

For sweep: Read summary.json. Extract:

• total_runs, total_hypotheses
• n_bonferroni_significant or bonferroni_survivors
• swept_parameters
• Top 3 most significant results (by effect size)
• p_hacking_audit

For redteam: Read report.json. Extract:

• robustness_score, grade
• attacks_tested, attacks_successful
• vulnerabilities (critical and high severity)

For study: Read analysis/summary.json or summary.json. Extract:

• descriptive statistics per condition
• pairwise_tests — significant comparisons
• bonferroni_survivors

For single: Read history.json. Extract:

• Final epoch metrics: welfare, toxicity, acceptance rate
• Agent-type breakdown if available