swarm: Skill for Claude Code

EXECUTE NOW

Target: $ARGUMENTS

Parse immediately:

• If target is a path: use it as the vault root
• If target is empty: use the current working directory
• If target contains --backfill: also generate run.yaml for existing runs
• If target contains --force: overwrite existing templates and index
• If target contains --minimal: skip seeded content notes, only install templates

Execute steps 0 through 8 in order. Do not skip steps. Do not ask for confirmation between steps — execute the full sequence, then report.

Step 0: Detect Vault State

Read the target directory to determine what already exists.

Check for:

• vault/ directory (has vault been initialized before?)
• vault/_index.md (has the index been created?)
• vault/templates/ (have templates been installed?)
• runs/ directory (are there existing runs?)
• schemas/ directory (have schemas been installed?)
• run-index.yaml (has the run index been built?)

Decision tree:

• If vault/_index.md exists AND --force not set → report "Vault already initialized" with a summary of what exists, then offer to run --backfill or --force
• If vault/_index.md does not exist → proceed with full initialization
• If --force is set → proceed, overwriting existing files

Store the state as context for remaining steps.

Step 1: Create Directory Structure

Create the full vault directory tree. Use mkdir -p to be idempotent.

vault/
├── experiments/          # one note per run (auto-generated)
├── claims/               # claim cards with evidence pointers
├── governance/           # governance mechanism notes
├── topologies/           # topology behavior notes
├── failures/             # named failure patterns from red-team
├── methods/              # reusable methodology notes
├── sweeps/               # cross-run sweep aggregations
└── templates/            # schema-enforced note templates

Also ensure these exist:

schemas/                  # JSON Schema files for validation
scripts/                  # automation tooling