workflow-skills

A bundled Claude Code plugin of multi-agent workflow skills plus a library of audit-lens subagents. The skills orchestrate fan-out/verify/loop patterns for research, verification, gap-finding, copywriting, implementation review, and a continuously-running codebase audit.

Install

/plugin marketplace add xdevs23/workflow-skills
/plugin install workflow-skills@workflow-skills

Then the skills appear in the skill list and each has a matching slash command (e.g. /research-loop, /audit-loop).

What's inside

Skills

Skill	What it does
research-loop	Triangulate an open research question into a trustworthy, evidence-tagged doc.
verify-loop	Adversarially prove every claim in an artifact against ground truth, loop to immaculate.
find-gaps	Audit an artifact for sins of omission — what's missing or under-specified.
implement-review-verify	Cold-review the spec first, implement, then four parallel verdict seats (correctness, cleanliness, spec compliance, duplication) plus two human-lane adversary seats. Every finding routes by defect class — mechanical to the fixer under re-verification rounds, taste and design-authority to the human — and the fix pass proves the suite green.
copywriting	Write an increment's user-visible strings BEFORE implementation: intent catalog + writing system, one agent per item, mechanical gate + source-verify + fresh-context critic, human ships the load-bearing lines.
immaculate-spec-writing	Research → find-gaps → verify convergence loop for a fully factual, complete spec.
resume-interrupted-run	Recover a workflow run that was stopped while agents were mid-flight: hand each interrupted seat its own prior transcript, leave every completed prompt byte-identical, resume near-losslessly.
audit-loop	Continuously audit a codebase through 8 lenses, append verified findings to AUDIT.md — plus a Refuted ledger of killed claims, so nothing is rediscovered every round.
group-chat	Join a shared multi-instance group chat so this Claude can talk to other Claude instances across machines/Docker, over a common hub. Built on Claude Code Channels.

Audit-lens subagents (sonnet-pinned, read-only)

separation-of-concerns, abstraction-quality, code-smell, type-safety, code-cleanliness, missing-gaps, domain-leakage, type-smearing. Usable directly as agentTypes in your own workflows, and used by audit-loop.

Requirements / assumptions

• The Workflow tool / multi-agent fan-out. Every skill orchestrates subagents via Workflow. A harness or plan that doesn't expose Workflow can't run these.
• Model tiers. Agents are pinned by convention: sonnet is the floor for review/research/audit lenses; opus for any coding/fix stage; never haiku. A Sonnet-only plan will not honor the opus pins; an opus-capable plan will incur opus rates on coding stages. This is deliberate — the model discipline is part of the product.
• audit-loop runs continuously. It's designed to be driven by /loop audit-loop in a tight loop with no interval — each ~30+ min full-tree round re-fires the next immediately. It is read-only and append-only to a project-local ./.claude/workflow-skills/AUDIT.md (commit or gitignore it as you prefer); it never fixes. Fix everything in one sweep when you choose (implement-review-verify pairs well).

Group chat (multi-instance) — setup

The group-chat skill lets several Claude Code instances — on different machines, in or out of Docker — join shared named groups and message each other. It is built on Claude Code Channels (research preview), so incoming messages push into the session continuously as <channel> events.

It has two pieces:

• The hub (servers/group-chat-hub/hub.ts) — one standalone process you run wherever you want the shared meeting point. It holds the per-group member list and an on-disk JSONL message log, and fans messages out. Run it with Bun:

  GROUP_CHAT_TOKEN=some-secret bun servers/group-chat-hub/hub.ts
  # listens on ws://127.0.0.1:8787 by default
  # env: GROUP_CHAT_PORT, GROUP_CHAT_HOST, GROUP_CHAT_DATA, GROUP_CHAT_ALLOW_NO_AUTH=1

• The channel adapter (group-chat/adapter.ts) — ships inside this plugin and is declared in the plugin's root .mcp.json. Claude Code spawns it per session. Point it at the hub with an env var (token inline in the URL, optional):

  export GROUP_CHAT_URL="ws://some-secret@your-hub-host:8787"

  Then launch Claude Code with the channel enabled (custom channels need the dev flag during the research preview):

  claude --dangerously-load-development-channels plugin:workflow-skills@workflow-skills

Once running, load the group-chat skill and join a group. Messages from other instances arrive as <channel source="group-chat" group="..." from="..." ...> events; you send with submit_message. Requires claude.ai (Pro/Max) or Console API-key auth — Channels are not available on Bedrock/Vertex/Foundry.

License

MIT