Claude Code Pipeline Kit

A drop-in governance harness for Claude Code: deterministic hooks, per-run pipeline state, and a complete SE/EIW/DRW/PDCA workflow library. Ships as a self-contained .claude/ directory.

Structured engineering pipelines for Claude Code. Multi-phase quality gates, persona-based reviews, deterministic hook enforcement, and a self-improving learning cycle -- in a single drop-in .claude/ directory.

This kit transforms Claude Code from a conversational assistant into a governed engineering harness:

• 9 deterministic hooks enforce CLAUDE.md MUST rules at the right tool boundary (output gating, phase gating, model enforcement, lint, migration reminder, completion enforcement).
• 4 pipelines for the full software lifecycle: full-cycle (SE), implementation-only (EIW), defect resolution (DRW), and self-improvement (PDCA).
• 18 procedural skills + 9 agent personas + 2 rule sets keep every stage anchored to the same standards.
• Pipeline state persistence survives compaction and session restart so long-running pipelines do not lose progress.
• English-only, vendor-neutral, and license-clean (MIT).

---

Quickstart

git clone https://github.com/your-org/claude-code-pipeline-kit
cd claude-code-pipeline-kit
bash install.sh --all

What install.sh --all does:

1. Copies .claude/skills/, .claude/agents/, .claude/rules/, .claude/commands/ into your project.
2. Compiles the 9 hooks (npm install && npm run build) into .claude/hooks/dist/.
3. Copies .claude/pipeline-state/SCHEMA.md so the state mechanism is documented in your repo.

Then enable the hooks by copying the settings template:

cp .claude/settings.json.template .claude/settings.json

Restart Claude Code. The next /se-pipeline, /eiw-review, /defect-fix, or /pdca-cycle invocation will run end-to-end with full hook enforcement.

First time? Read docs/quickstart.md for the 5-minute walk-through.

---

What's inside

Production-ready governance for AI-assisted development. Drop the kit into your repo to get the deterministic enforcement and per-run state layer that powers the SE/EIW/DRW/PDCA pipelines.

Component	Count	Location
Skills (procedural workflow guides)	18	.claude/skills/
Agents (persona reference cards)	9	.claude/agents/
Rules (coding standards)	2	.claude/rules/
Commands (pipeline + utility skills)	40	.claude/commands/
Hooks (TypeScript, compiled to dist/)	9	.claude/hooks/

Pipelines:

• SE Pipeline -- Full lifecycle (Phase 0 Exploration -> Phase 9 CEO approval), 12 phases including UX Bar Raisers at 5.5 and 7.5.
• EIW (Enterprise Implementation Workflow) -- 8 stages for implementation-only tasks (Stage 0 Architecture Review -> Stage 7 CEO approval) with Stage 3.5 Bar Raiser.
• DRW (Defect Resolution Workflow) -- 6 stages for bug fixes (D1 Investigation -> D5 Technical Review) with D3.5 Bar Raiser.
• PDCA -- 4 phases of autonomous self-improvement (Incident -> Attribution -> Synthesis -> Upgrade) triggered automatically after every error or feedback resolution.

Hooks (deterministic CLAUDE.md enforcement):

Hook	Lifecycle	What it gates
pretool.gate-skill	PreToolUse(Skill)	Phase-correct skill invocation
pretool.gate-output	PreToolUse(Edit/Write)	File writes outside .claude/ only during implementation phases
pretool.enforce-opus	PreToolUse(Task)	Mandatory Opus model on subagents
posttool.lint-ui	PostToolUse(Edit/Write)	Optional UI vocabulary check
posttool.lint-skill	PostToolUse(Edit/Write)	Skill-file lint (front-matter, structure)
posttool.remind-migration	PostToolUse(Write)	Optional migration reminder
session.resume	SessionStart	Detect resumable pipelines + dist freshness
stop.flush-manifest	Stop	Persist run manifest atomically
stop.enforce-pipeline-completion	Stop	Block premature stop while a pipeline is mid-flight

Full reference: docs/hooks.md. Customisation: docs/customization.md.

---

Architecture

The kit follows a CLAUDE.md > skills/agents/rules > instincts precedence model:

        ┌─────────────────────────────────────────┐
        │  CLAUDE.md  (governance, principles)    │
        └────────────────┬────────────────────────┘
                         │
         ┌───────────────┼───────────────┬─────────────┐
         ▼               ▼               ▼             ▼
   .claude/skills/   .claude/agents/   .claude/rules/   .claude/commands/
   (workflow)        (personas)        (style)          (pipelines)
                         │
                         ▼
               .claude/hooks/  (deterministic enforcement)
                         │
                         ▼
            .claude/pipeline-state/  (per-run state)
                         │
                         ▼
            .claude/pdca-archive/    (learning history)

Pipelines run through a sequence of Step A Discussion -> Step B Convergence -> Step C Deliverable -> Step D Approval within each phase. Bar Raisers inject between major implementation gates and force one mandatory UX redo per pipeline.

Full overview: docs/architecture.md.

---

Pipelines at a glance

Intent classification (4-tier)

The kit installs a 4-tier decision tree into your CLAUDE.md so Claude Code automatically routes work:

User intent	Pipeline	Trigger
Read-only / advice	None	Conversational
Trivial fix (1 file, ≤3 lines, cosmetic)	None (with sentinel)	Apply directly
Bug / error / test failure	/defect-fix (DRW)	Investigation -> scope -> TDD fix -> verification -> review
New feature / artifact / undefined design	/se-pipeline (SE)	Full 10-phase lifecycle
Defined design, implementation only	/eiw-review (EIW)	8 stages, lighter gate set

Full pipeline reference: docs/pipelines.md.

---

Customisation

The kit ships sensible defaults but every gate is configurable:

• Model override -- swap model: "opus" for model: "sonnet" to reduce cost.
• UI lint vocabulary -- enable posttool.lint-ui and add brand-specific terms.
• Migration reminder -- enable posttool.remind-migration and supply your project's migration path regex.
• Custom skills -- author project-specific workflow skills under .claude/skills/.

Full guide: docs/customization.md.

---

Documentation map

docs/
├── README.md            # Reading order for newcomers
├── quickstart.md        # 5-minute first run
├── architecture.md      # Topology, lifecycle, data flow
├── hooks.md             # Per-hook reference
├── pipelines.md         # SE / EIW / DRW / PDCA full reference
├── customization.md     # Model, lint, custom skills
├── contributing.md      # Repo structure, dev workflow, tests
└── settings-reference.md # Annotations for .claude/settings.json.template

---

Requirements

• Claude Code (latest)
• Node.js 20+ (the hooks compile from TypeScript -> ESM)
• npm (ships with Node)
• bash 3.2+ (the install script is bash 3.2-compatible — runs on stock macOS)

---

License

MIT -- use it commercially, modify it, redistribute it. The kit is offered as-is with no warranty.

---

Acknowledgements

This kit is a generalised port of an internal pipeline harness used to build a production AI engineering platform. It is contributed under MIT for general use.