pipeline

Pipeline

Structured, multi-agent software development for any AI coding tool.
From agentic development to agents as the developer team.

---

Freeform "vibe coding" with an agent fails at scale: no separation between deciding what to build and building it, the author grades their own homework, scope creeps, review gets skipped when "it's simple," and nothing compounds.

Pipeline replaces that with structure. Work is broken into work packages — small, outcome-level specs with acceptance criteria — and each runs through a fixed phase loop driven by three separated personas:

work package ──▶ design ──▶ critique ──▶ build (TDD) ──▶ review ──▶ retro ──▶ ship
                planner    reviewer      builder        reviewer    fresh     builder

• The agent that designs is not the agent that reviews it. Producer/evaluator separation is enforced by persona.
• Planning is phase 1, never the finish line. A plan isn't done until the builder makes it real and the reviewer signs off.
• Gates are mechanical. Your verify command must pass and the review verdict must be DONE before ship.
• Phases that don't apply are skipped — a backend work package skips the design phases automatically.

What's in here

• skills/ — the pipeline skills, each a SKILL.md (the Agent Skills open standard).
• agents/ — planner / reviewer / builder persona subagents.
• hooks/ — a session-start hook (surfaces the pipeline) and an edit-streak hook (nudges the orchestrator to delegate after 5 consecutive edits). Both are wired for Claude, Cursor, Gemini, Codex, and Copilot via .cursor/, .codex/, .gemini/, .github/. The edit-streak skips subagent edits on Claude (which exposes that distinction); on the others it's best-effort.

Install

Claude Code — plugin

/plugin marketplace add ambrovia/agent-skills-pipeline
/plugin install pipeline@agent-pipeline

Skills become /pipeline:concept, /pipeline:review, … and the orchestrator /pipeline.

Cursor · Copilot · Gemini · Codex — copy the files

These tools read SKILL.md from their own directory. Copy skills/ (and agents/) into it, plus hooks/ and your tool's hook config (.cursor/, .gemini/, .codex/, or .github/) so the pipeline surfaces at session start:

Tool	Put skills in
Cursor	.cursor/skills/ or .agents/skills/
Copilot	.github/skills/ or .agents/skills/
Gemini / Antigravity	.gemini/skills/ or .agents/skills/
Codex	.agents/skills/

.agents/skills/ is the shared standard for all of these. Claude Code is the exception — it reads .claude/skills/, so use the plugin (or copy into .claude/).

Configure

Everything project-specific lives in one file. Copy pipeline.config.example.yml to pipeline.config.yml; skills resolve {{key}} from it:

verify: "go test ./..."   # the single command that must pass before ship
designSystem: null        # null → the design phases are skipped
vcs: github

The skills

concept · design · architecture · design-critique · architecture-critique · write-tests · write-code · write-docs · review · ship · retro · compound · lore · work-planning · pipeline

Run a whole work package through every applicable phase with /pipeline <id>. After several work packages, run /compound to mine the retro log for recurring patterns and propose process fixes. Use /lore anytime to capture or surface tribal knowledge.

License

Apache-2.0.