get-shit-done: Context Engineering

gsd-build/get-shit-done (64K stars, 2 weeks). The whole repo is a context engineering framework — meta-prompts that make Claude Code dramatically more effective.

The 3 layers

Layer 1: Project Context

The agent doesn’t know your project. Tell it:

# /AGENTS.md
This is a Next.js 15 SaaS app.
- Auth: Clerk
- DB: Postgres via Drizzle ORM
- Payments: Stripe
- Deployment: Vercel

Conventions:
- Use server actions, not API routes
- All forms use react-hook-form + zod
- Tests in __tests__/

Layer 2: Task Context

Each task gets its own brief:

# Task: Add CSV export to /reports
- Read /app/reports/page.tsx
- Add a "Export" button
- Use papaparse
- Stream to client (no full page reload)
- Test: 10k row export in <5s

Layer 3: Execution Context

What the agent should do RIGHT NOW:

You are working on Task #1 above.
- Do NOT touch unrelated files
- Do NOT refactor existing code
- Run `npm test` after every change
- Commit when done with message: "feat(reports): CSV export"

The “Done” pattern

Every task ends with a checklist:

When done, write to /CHANGELOG.md:
- [ ] Tests added
- [ ] Tests pass
- [ ] No console.log
- [ ] Migration added (if schema changed)
- [ ] Self-review: any TypeScript errors?

Why it works

Most AI coding failures are context failures, not model failures. The agent doesn’t know:

• Your project’s hidden rules
• What “done” means for you
• What NOT to touch

get-shit-done makes context explicit.