Introducing Claude Code

Your AI Development Team

AI agents that write, test, review, and deploy production code. Not autocomplete — autonomous development.

Sessions

Commits

Completion Rate

Building Blocks

Three pillars that transform Claude Code into a development workforce.

CLAUDE.md

The constitution

Project-level instructions. Critical rules, coding standards, skill registry, agent definitions.

# Claude Configuration

## TL;DR

- Before any work: Read dev-process.mdx

- All code changes require a worktree

- Always use Conventional Commits

16 Skills

Slash commands

Slash-invoked prompts that teach Claude Code complex, multi-step tasks with consistent quality.

/review/bootstrap/scaffold/pr/promote/test+10 more

9 Agent Definitions

.claude/agents/

Role-specific .md files defining tools, permissions, domains, and skills.

frontend-devbackend-devdevopstesterfixersecurity-auditorarchitectproduct-leaddoc-writer

Combine these three and you get something powerful — a specialist.

The Specialization Formula

A generalist agent produces generic UI. A specialist produces production quality.

agent.md

Role definition

Scoped file boundaries, permitted tools, domain permissions. The agent knows exactly what it owns.

apps/web, packages/ui

Docs

Standards & patterns

Mandatory reads before every task. Frontend patterns, testing standards, code review checklists.

frontend-patterns.mdx, testing.mdx

Skills

Slash commands

Reusable skills that extend capabilities. Design guidance, library docs, commit automation.

ui-ux-pro-max, context7

Specialist

The output

A focused agent that produces production-quality output instead of generic boilerplate.

FE dev with design sense

Example: frontend-dev agent

Knows its file boundaries (apps/web, packages/ui), reads frontend-patterns.mdx before every task, and has the ui-ux-pro-max skill for design decisions. A generalist agent produces generic UI. A specialist produces production quality.

Development Process

A 50-file mechanical change may be F-lite. A 3-file rate limiter may be F-full.

Match process to complexity

Tier SQuick Fix

3 files or fewer, no architecture risk

Worktree + PR

Tier F-liteFeature (lite)

Clear scope, single domain

Worktree + agents + /review

Tier F-fullFeature (full)

New architecture, unclear requirements, >2 domains

Bootstrap + worktree + agents + /review

Key insight: A 50-file mechanical change may be F-lite. A 3-file rate limiter may be F-full.

Worktree = Parallelism

3 features simultaneously

Each issue gets its own worktree — isolated branch, deps, and DB. One Claude Code session per pane.

#42 -- Auth module

> /scaffold --issue 42

#43 -- API endpoint

> /scaffold --issue 43

#44 -- Dashboard

> /scaffold --issue 44

Quality Gates

Git Hooks (Lefthook)

pre-commitBiome lint + auto-fix staged files

commit-msgCommitlint validates Conventional Commits

pre-pushFull lint, typecheck, test coverage

Claude Hooks

PostToolUseBiome auto-format on every file edit

PreToolUseSecurity warnings before file changes

Rule: Lefthook for hard gates. Claude hooks for DX. Claude can't bypass git hooks.

The Toolchain

16 skills organized by workflow phase. Each skill is a reusable, multi-step prompt.

Plan

Requirements, specs, triage, decisions

/interview/bootstrap/issue-triage/adr

Build

Code from specs, test generation

/scaffold/test

Review

Multi-agent review, finding walkthrough

/review/1b1

Ship

Pull requests, releases, cleanup

/pr/promote/cleanup

Maintain

Quality, retrospectives, monitoring

/validate/retro/issues/agent-browser/compress

Plan

Requirements, specs, triage, decisions

/interview/bootstrap/issue-triage/adr

Build

Code from specs, test generation

/scaffold/test

Review

Multi-agent review, finding walkthrough

/review/1b1

Ship

Pull requests, releases, cleanup

/pr/promote/cleanup

Maintain

Quality, retrospectives, monitoring

/validate/retro/issues/agent-browser/compress

Agent Teams in Action

9 agents, 3 categories. The orchestrator delegates — it never writes code.

orchestrator

Spawns agents, assigns tasks, synthesizes. Never writes code — preserves context window.

Spawn specialized agents via Task tool

Assign tasks, track dependencies

Synthesize results, report to human

Write or edit code directly

Write documentation content

Run tests or linters

Strategy

Planning & design

architectproduct-leaddoc-writer

Domain

Code writers

frontend-devbackend-devdevops

Quality

Verification

testerfixersecurity-auditor

Review Pipeline

Fresh agents review (never the author)

Conventional Comments per domain

Human decides each finding via AskUserQuestion

Parallel fixers apply accepted fixes

Why fresh agents? An agent that wrote the code has sunk-cost bias — it won't question its own decisions. A fresh reviewer sees the code for the first time and challenges every choice, like a real code review.

Idea to Production

Analysis, spec, code, review, deploy. Claude manages the full SDLC — including GitHub and Vercel.

/interview

Idea → Interview

/interview: structured requirements gathering

/bootstrap

Interview → Analysis

Interview, structured analysis

validation gate

/bootstrap

Analysis → Spec

Analysis promoted to spec

validation gate

/scaffold

Spec → Code

Plan approval, then agents implement

/pr

Code → PR

Create pull request

/review

PR → Reviewed

Fresh review + /1b1 decisions

validation gate

Merge

Reviewed → Staging

PR merged into staging branch

/promote

Staging → Main

Preview verification, then deploy

validation gate

/interviewIdea → Interview

/interview: structured requirements gathering

/bootstrapInterview → Analysisgate

Interview, structured analysis

/bootstrapAnalysis → Specgate

Analysis promoted to spec

/scaffoldSpec → Code

Plan approval, then agents implement

/prCode → PR

Create pull request

/reviewPR → Reviewedgate

Fresh review + /1b1 decisions

MergeReviewed → Staging

PR merged into staging branch

/promoteStaging → Maingate

Preview verification, then deploy

GitHub

Issues, Projects, PRs, Actions

Vercel

Deploy, logs, env vars, rollback

CI/CD

GitHub Actions + Vercel auto-deploy

Lessons Learned

Honest insights from 1,930 sessions of AI-assisted development.

What works

Subagents via Task tool — Claude handles coordination internally
Specialized agents with domain docs + skills
Fresh reviewers catch what authors miss

What doesn't

Multi-panel Agent Teams — coordination overhead exceeds benefit
Orchestrator writing code — burns context, produces worse results
Generalist agents — produce generic, boilerplate output

What changed

Bootstrap evolved: interview → analysis (gate) → spec (gate)
Skill ecosystem grew: 0 → 16 custom skills in 25 days
Human gates at every decision — not just code review

Get Started

Everything you need to replicate this process.

Explore the Docs

Fork the Repo

Try /bootstrap on your first feature

Entirely built with Claude Code.