AI Builder Club — Skills

A Claude Code plugin marketplace of the skills we share at AI Builder Club for building loop engineers: agents that get triggered on their own, pick up work, ship it, verify it, and log what they learned, so the work compounds without you prompting every step. It's the productized version of the setup my team runs in production.

Two flagship skill sets (more to come):

• Codebase harness — make any repo agent-ready (run, test, verify, ship — including an isolated cloud box per agent so loops ship code in parallel).
• Loops — spin up compounding agent loops on a shared, file-based knowledge base.

Plus standalone utility skills:

• agent-context-audit — audit your agent context (CLAUDE.md, codebase docs, skills, tool/MCP designs) against Anthropic's Claude 5 context-engineering guidance — finds overconstraint, conflicting instructions, redundancy, stale facts, and missing gotchas; reports findings with concrete rewrites, then applies approved fixes.
• Open agent teams — delegate tasks to any CLI agent (claude, codex, grok, aider, …) running in a detached tmux session, with a race-safe done-signal protocol, multi-turn iteration, and a CLAUDE.md delegation-rules template (coordinator/executor roles).
• SEO growth — where to point SEO effort, as two playbooks: a cold start with no rankings and no authority, and an existing site with data. Distilled from two content properties we run in production — including the emerging-term method that took a brand-new pillar page to the site's #1 clicked page in 7 days, and the failures that cost us a head term. Ships a loop template so it can run on a schedule.
• Visual flow GIF — turn an article, workflow, or architecture into a static PNG + animated GIF flow diagram (JSON spec → local Python/Pillow renderer).
• Karaoke captions — burn word-level highlight captions into a video (MLX Whisper → ASS → FFmpeg libass).

Loop engineer & Codebase harness

The shift: you stop prompting a coding agent task-by-task, and start designing loops.

A loop is an agent that wakes up on a trigger (a cron, a webhook, an incident, another agent), does some investigation and work, and writes what it found and did into a shared, file-based memory. Next run it reads that memory and keeps going. The real power is compounding: many loops (support, SEO, product, ads) read and write the same folders, so a friction the support loop logs can get picked up by the product loop, and a keyword the ads loop finds can feed the SEO loop. One shared brain, many loops.

Building one comes down to four ingredients:

1. Triggers: cron, webhook, an incident, or another agent wakes the loop at the right time.
2. A file + logging structure: the shared memory loops read and write → the loops skills.
3. Tools & connectors: so the agent can do real work (your skills/MCPs).
4. A codebase harness: so the agent can run, test, and verify its own work → the codebase-harness skills.

This single skills plugin ships both sets — giving you #2 and #4, plus the scaffolding to add the rest.

Want the full walkthrough of the concept and how my team designs compounding loops? Watch the video:

Install

/plugin marketplace add AI-Builder-Club/skills
/plugin install skills@ai-builder-club

One plugin, all the skills below.

The two entry points

/setup-codebase-harness — make a code repo agent-ready

Run it in the code repo your agents work in, so they can run, test, and verify their own work. It orchestrates the harness skills below — pull in only what the repo needs.

/new-loop — build your shared brain

Run it where your agent's memory should live. First run bootstraps the knowledge base (creates ARCHITECTURE.md, LOG.md, the signals/ docs/ domains/ folders, and a knowledge-base section in your CLAUDE.md); then it scaffolds the loop, does one real test run, and logs it. Run it again any time to add another loop.

The skills (when to use which)

Codebase harness — make a repo agent-ready

Loops — the shared knowledge base

Visuals — turn source material into diagrams

Video — burn karaoke-style captions into a video

Context — keep your agent context sharp

Delegation — run other agents as executors

After setup, each session the agent reads CLAUDE.md + the relevant domain README, does work, writes artifacts, and appends to LOG.md. For code changes it works in an isolated worktree and verifies via /verify before shipping.

Requirements

• Claude Code (the skills assume it).
• git. That's the only hard dependency for the knowledge base + harness.
• The harness skills want the code repo they ship into to be a git repo with a working build/test setup. They use Codex for review if available, and degrade gracefully if not.
• crabbox-setup (optional, for parallel cloud boxes) needs the crabbox CLI + a provider (Daytona: daytona CLI / DAYTONA_API_KEY).
• open-agent-teams needs tmux, plus whichever CLI agents you delegate to on your PATH (claude, codex, grok, pi, opencode, …).
• karaoke-captions (optional) needs macOS Apple Silicon, Python 3.10+, and FFmpeg with libass.

Repo layout

skills/                                a Claude Code plugin (also a marketplace)
├── .claude-plugin/
│   ├── marketplace.json              marketplace: ai-builder-club
│   └── plugin.json                   plugin: skills   (source ".")
└── skills/                           top-level skills
    ├── new-loop/                     (loops) — + references/: ARCHITECTURE · LOG · KNOWLEDGE_SETUP · CLAUDE.template
    ├── setup-codebase-harness/       (harness) — orchestrator
    ├── dev-local-setup/              (harness)
    ├── e2e-setup/                    (harness)
    ├── crabbox-setup/                (harness) — isolated cloud box per agent
    ├── verifier-setup/               (harness) — generates a repo's /verify skill  (assets/verify.template.md)
    ├── visual-flow-gif/              (visuals) — JSON spec → PNG + GIF  (scripts/ assets/ references/)
    ├── agent-context-audit/          (context) — audit CLAUDE.md/docs/skills/tools vs Claude 5 guidance
    ├── seo-growth/                   (growth) — cold-start + existing-site SEO playbooks