Stop Letting AI Write Code You Haven't Approved

I've been building AI solutions using spec-driven development for 2 years. Along the way I packaged the workflows into attune-ai, an open-source plugin for Claude Code. The AI stays aware of your spec and where you are in the implementation process. You still do the real engineering work -- but your starting prototypes come out higher quality from day one.

---

/spec: Brainstorm, Plan, Review, Then Execute

Attune AI v5.3 adds a /spec command to Claude Code that enforces a simple lifecycle:

--- CODE START ---
/spec add rate limiting to the API
--- CODE END ---

What happens:

1. It ASKS questions first (what's the problem? what does success look like?)
2. It decomposes the work into discrete tasks with files, validation checks, and risks
3. You REVIEW the full plan before any code is written
4. It executes task-by-task with quality gates after each

If a quality gate finds a security issue, it blocks. You can't auto-approve high-severity findings.

If you close your laptop mid-spec, it remembers where you left off. State lives in the plan file itself.

---

Why This Matters

The spec file is just markdown with XML tasks -- version controlled, portable, human-readable.

--- CODE START ---
Status | ID | Name            | Objective
done   | 1  | rate-limiter    | Core sliding window
>>>    | 2  | middleware      | FastAPI integration
...    | 3  | redis-backend   | Distributed support
--- CODE END ---

This is how experienced engineers already work -- scope the work, review the plan, then execute. /spec just makes AI follow the same discipline.

---

Try It

--- CODE START ---
pip install 'attune-ai[developer]'
attune setup
--- CODE END ---

Then type /spec in Claude Code.

Open source, Apache 2.0:
github.com/Smart-AI-Memory/attune-ai

#ClaudeCode #AIEngineering #DeveloperTools #OpenSource
