A governed, phased process for building with Claude Code: tiered pipelines, STOP gates you approve, audit skills, and a human-approval commit gate. Built on Claude Code primitives.

Claude Code-only. Tierward's enforcement layer (STOP gates, the human-approval commit gate, and audit skills) is built on Claude Code primitives and does not support other AI coding tools. See why in the docs.

npx tierward init

What it does

Claude Code is fast. The gap it creates is not in the code; it is in the review. When an AI agent writes autonomously, decisions accumulate faster than anyone can verify them. Tierward closes that gap by enforcing a review contract directly in the development session, not as a prompt Claude may or may not follow, but as a process it cannot bypass.

You stay the decision-maker. STOP gates stop Claude before implementation: it presents its plan and waits for your explicit approval before writing code. A commit gate then blocks git commit until that approval is on record, and CODEOWNERS prevents anyone from modifying .claude/ without a human reviewer.

Verification is built in at the same level. The Stop hook is a shell command in .claude/settings.json: when Claude tries to declare a task done, it runs your test suite, and if tests fail Claude is blocked and keeps working. It can't skip this step.

Quick start

# Add Tierward to an existing project
npx tierward init

# Validate the scaffold
npx tierward doctor

# Check for Anthropic spec drift
npx tierward upgrade --anthropic --dry-run

The wizard detects your tech stack and scaffolds the right configuration. Runs in about two minutes.

Four tiers

Start at the lowest tier that covers your risk. Move up when you need more structure.

Tier	Pipeline	Best for
0 — Discovery	Stop hook only	First exploration, zero process
S — Fast Lane	4 steps, scope-confirm	Solo dev, low-risk fixes
M — Standard	13 phases, 3 STOP gates	Feature blocks, 1–2 collaborators
L — Full	14 phases, 4 STOP gates	Team projects, complex domain work

What gets scaffolded

CLAUDE.md — project context Claude reads at session start (stack, commands, conventions)
.claude/settings.json — Stop hook, permissions allow/deny list, audit log hooks
.claude/rules/pipeline.md — the development pipeline Claude follows, phase by phase
.claude/rules/security.md — stack-aware security rules (11 stacks supported)
.claude/skills/ — audit slash-commands: /arch-audit, /security-audit, /systematic-debugging, and more
.github/CODEOWNERS — gates .claude/ changes behind human review
team-settings.json (opt-in) — enforce minTier, allowedSkills, blockedSkills across every team clone

Audit skills

26 executable multi-step programs that run inside Claude Code. Not one-shot prompts: structured workflows with STOP gates, model routing (haiku for mechanical checks, sonnet for analysis), and structured output.

/security-audit /arch-audit /doc-audit /test-audit /accessibility-audit /compliance-audit /api-contract-audit /infra-audit /dependency-audit /perf-audit /visual-audit /responsive-audit /ux-audit /migration-audit /pr-review /systematic-debugging /simplify and more.

MCP server

npx -y --package=tierward tierward-mcp

Or install from the MCP registry: io.github.marcoguillermaz/tierward

Six read-only tools expose governance state to any MCP-aware client (doctor report, team settings, arch-audit status, skill inventory, package metadata) without the CLI running.