Slow Claude — Development skill for Claude Code

slow-claude

Slowing Claude down is the feature, not the bug.

A Claude.ai skill + Claude Code slash commands that enforce a disciplined feature development workflow. Stops Claude from jumping straight into code, demands evidence before recommendations, and uses Spec Driven Development with explicit approval gates between phases.

Built for solo founders shipping a feature a day under competitive pressure.

---

The problem

You're moving fast. A new feature comes up. You open Claude and say "I want to add X."

What happens by default:

• Claude dumps code based on memory or vibes
• Recommends a library it half-remembers from a blog
• Skips the questions that would have changed the answer
• Lets the coding agent (Claude Code, Cursor, etc.) start implementing before you've decided the right approach

Three days later you realize the library is wrong, the persistence model is wrong, or the edge cases weren't considered. You rewrite. At one-feature-a-day cadence, this rework is what actually slows you down.

The fix

**slow-claude** turns Claude into a structured architect. Not faster — more deliberate. The workflow is 4 phases, each with an approval gate:

1. Research — Claude Code creates a feature/ branch, reads your codebase, compares options with citations (GitHub issue numbers, measured bundle sizes), and writes RESEARCH-.md. ~10 minutes.
2. Decision — You read the research, approve a choice, answer open questions.
3. SDD Build — Zod schemas first. Then phases A→F with build-clean gates between them. Migrations generated but never auto-applied.
4. Verify — Manual test against real DB and real auth (demo mode bypasses persistence). Open PR.

The skill lives in Claude.ai. The slash commands live in Claude Code. Templates are shared between them — single source of truth.

---

What's in this repo

slow-claude/
├── README.md                     ← you are here
├── feature-dev-workflow/         ← the Claude.ai skil