← All experiments
Agent SkillsOn GitHub

Muster

Structured agentic development as Agent Skills. Seven commands take a half-formed idea to a merged pull request, with sub-agents doing the building and a person approving every gate.

Details

Started
22 Jun 2026
Commands
think, context, spec, triage, build, review, status
Runs in
Claude Code, Cursor, Codex, Gemini CLI, Mistral, pi
Merging
Always a person, never automatic
Format
Agent Skills, no tool lock-in
Language
Shell
Licence
MIT
Last commit
6 Nov 2026

ConjectureMost agent failures come from requirements nobody pinned down, not from bad code.

FindingStill open.

The write-up

Most agent failures are not the model writing bad code. They are requirements that were never pinned down. Muster is built on that observation, and on one more borrowed from disciplined engineering: separate deciding what to build from building it.

Seven commands, in order

  1. /think asks questions until the idea is clear.
  2. /context writes down how the project works, once per repo.
  3. /spec turns the idea into a short plan and a set of small task issues on GitHub.
  4. /triage is the Manager: it sorts the tasks, judges the risk and marks the safe ones ready.
  5. /build sends out the builders, sub-agents that build each task test-first, check their own work and wait for approval before opening a pull request.
  6. /review checks a change two ways, against the project’s conventions and against the issue it claims to solve.
  7. You merge. Nothing is ever merged automatically.

/status gives a plain summary of where everything stands at any point.

Why Agent Skills

Muster ships as plain Agent Skills rather than as a plugin for one tool, so the same workflow runs in whichever harness you happen to be using. Where a harness cannot fan out parallel sub-agents, /build degrades gracefully and does the work in sequence.