The Wizard of AI
Back to Resources
Agent Skill Pack

Types ofLoops

Four complete autonomous-loop packages for Claude Code — slash commands plus the specialized agents that power them. This is how you make an AI agent improve its own work instead of stopping at the first draft.

Download the Pack(.zip, 37 KB)

Loop 1 — Learning Loop (/skill-loop)

Autonomously iterates a skill-improver agent on any skill until it converges: exercise the skill on a representative task in a fresh context, record what went wrong in a LEARNINGS journal, apply one focused edit, repeat. Your skills get measurably better while you sleep.

Loop 2 — Multi-Agent Loop (/orchestrate)

Four independent critics — correctness, style, safety, and domain-expert — review a piece of content in parallel, findings get merged, the content gets revised, and the cycle repeats until all four sign off. Beats a single self-review every time.

Loop 3 — Verification Loop (/review-loop)

The heavyweight: an implementor agent builds against your PRD while a 'thermonuclear reviewer' runs max-intensity adversarial review — multi-lens fan-out, deduplication, multi-vote refutation, loop-until-dry. Includes a runnable Workflow script. Token-hungry by design; use it when correctness actually matters.

  • Requires a docs/prd.md and docs/rubric.md in the target repo
  • Findings are adversarially verified before they're reported — no plausible-but-wrong noise
  • Produces a mechanically-gated 0–100 score plus ranked, confirmed findings

Loop 4 — Workflow-Improving Loop (/iterate)

Kaizen for your process itself: a builder executes the current documented process, a scorer grades the output, and a process-optimizer makes ONE bounded change to the process doc per round with a falsifiable prediction. The process improves — not just the artifact.

How to Install

  1. 1Download and unzip the pack.
  2. 2Copy each package's commands/*.md into .claude/commands/ and agents/*.md into .claude/agents/ (project-level, or ~/.claude/ for global).
  3. 3Copy verification-loop/workflows/thermonuclear-review.js into .claude/workflows/ if you want the Workflow-engine version.
  4. 4Run /skill-loop, /orchestrate, /review-loop, or /iterate from Claude Code.

Pairs With