agentic-crew

Program management for AI agents. 12 specialized agents with defined roles, adversarial review loops, and automated quality gates -- organized the way a TPM organizes an engineering team.

---

Why This Exists

Most Claude Code setups are a single CLAUDE.md file with a few instructions. That's like running a project with one generalist and no process.

agentic-crew is a full multi-agent orchestration system -- specialized roles that debate decisions, challenge each other's output, catch bugs before they commit, and enforce consistency. Each agent has a defined scope, handoff protocol, and quality bar -- the same structure a program manager builds for human teams, applied to AI agents.

---

At a Glance

	Typical Claude Setup	agentic-crew
Agents	1 general assistant	12 specialized roles
Design review	Manual, single-pass	Adversarial -- critic challenges designer
Security	None	Threat modeling on every auth/API change
Pre-commit	None	Auto-blocks hardcoded secrets, debug logs, empty catches
AI feature work	Generic prompts	Dedicated ai-architect for LLM-native design
Code quality	Vibes	DRY checks, naming review, consistency audit

---

Setup

Symlink into ~/.claude/ so changes sync automatically:

# Option 1: Symlink (recommended)
ln -sf "$(pwd)/agents" ~/.claude/agents
ln -sf "$(pwd)/skills" ~/.claude/skills
ln -sf "$(pwd)/rules" ~/.claude/rules
ln -sf "$(pwd)/hooks" ~/.claude/hooks
cp CLAUDE.md ~/.claude/CLAUDE.md
cp settings.json ~/.claude/settings.json

# Option 2: Copy everything
cp -r agents skills rules hooks CLAUDE.md settings.json ~/.claude/

---

Agents (12)

The core of agentic-crew. Each agent has a focused role -- no generalists.

Agent	Purpose	Model
orchestrator	Coordinates multi-agent debate for 3+ file changes	default
design-thinking	Visual hierarchy and craft -- runs before UI work	default
design-critic	Adversarial review -- runs after UI code is written	default
designer	States, mobile, accessibility audit	default
code-reviewer	Consistency, DRY, naming review	default
qa-engineer	Edge cases and failure modes	default
security-engineer	Threat modeling for auth, data, APIs	default
product-manager	Validates problem and scope before new features	haiku
sre	Performance and reliability for backend work	default
ai-architect	AI-native design for LLM features	default
ux-copywriter	Copy quality and terminology	haiku
analytics-engineer	Data pipeline and metrics design	haiku

How the adversarial loop works

For UI tasks, agentic-crew runs a two-pass review:

design-thinking → [your code] → design-critic
     ↑                                 ↓
   revise  ←←←←←←←←←←←←←← critique + score

design-critic doesn't just suggest -- it challenges. You get a scored review with specific objections before anything ships.

---

Skills (7)

Slash commands that trigger structured, multi-step workflows.

Skill	Trigger	What It Does
/audit	Manual	Full codebase audit
/design-review	Manual	UI/UX design review
/review	Manual	Code review
/add-feature	Manual	Guided feature implementation
/fix-bug	Manual	Structured bug fix workflow
/refactor	Manual	Safe refactoring with tests
/red-team	Manual	Security red team assessment

---

Rules (6)

Language and domain-specific guardrails enforced across every project.

Rule	Covers
backend.md	API design, database, error handling, performance
frontend.md	React/Next.js, components, responsive, accessibility
git.md	Commits, branches, safety
python.md	Modern Python 3.10+, async, typing, anti-patterns
typescript.md	Strict mode, null safety, modern patterns
security.md	Input validation, auth, secrets, common attacks

---

Hooks

Automated checks that run at key moments -- no manual triggers needed.

Hook	Event	Type	What It Catches
lint-after-edit.sh	PostToolUse (Edit/Write)	command	Runs tsc or py_compile after every edit
Pre-commit check	PreToolUse (git commit)	prompt	Debug logs, hardcoded URLs, empty catches, secrets
Task review	TaskCompleted	prompt	Reviews output against original task description

---

Global Rules (CLAUDE.md)

Non-negotiables enforced across all projects:

• Think before coding -- plan first, implement second
• Mobile-first at 375px
• Descriptive error messages -- no generic "something went wrong"
• No dead code or duplicates
• TypeScript strict -- no any, ever
• All 5 UI states: loading, empty, error, partial, edge case
• Touch targets ≥ 44px, input font ≥ 16px
• Toast notifications -- never alert()

---

Who This Is For

• Developers shipping production apps with Claude Code
• Teams wanting consistent AI-assisted code quality without manual review overhead
• Anyone building LLM-powered features who needs an ai-architect perspective baked in

---

Contributing

Found a better agent prompt? A rule that should exist? PRs welcome.

Open an issue describing what gap you hit -- agent, skill, rule, or hook -- and let's build it.

---

License

MIT