This isn’t generic prompt engineering — These techniques are specific to working with Claude Code in real codebases. If you want general prompting advice, look elsewhere.
Be Specific, Not Emphatic
Error messages beat adjectives. File paths beat “the auth code.” ALL CAPS doesn’t help — context does.
| Weak | Strong |
|---|
| ”Fix the auth bug" | "The login endpoint returns 401 when the JWT is valid — see src/auth/middleware.ts:42" |
| "Make this FASTER" | "This query takes 3s on 10k rows. Profile it and suggest indexes." |
| "Write GOOD tests" | "Write tests covering the edge cases in processPayment: expired card, insufficient funds, duplicate charge.” |
Challenge Claude
Don’t just accept the first output. Push back:
- “Grill me on these changes and don’t make a PR until I pass your test.” — Forces Claude to verify you understand the code before shipping it.
- “Knowing everything you know now, scrap this and implement the elegant solution.” — After Claude has explored the problem space, ask it to restart with full context. The second attempt is almost always better.
- “What are three ways this could break in production?” — Surfaces failure modes you haven’t considered.
Pre-Planning: Reduce Assumptions First
Before jumping into a plan, ask Claude to research the codebase and surface 5-10 clarifying questions. This eliminates assumption drift — especially critical for multi-agent work where two agents assuming different things produces code that individually passes tests but fails integration.
Self-Reflection Loops
This is a Level 4 technique from the maturity ladder. Ask Claude to check its own work:
- “Are you sure about this? What about edge case X?”
- “Review this diff as if you were a senior engineer who’s skeptical of AI-generated code.”
- “What assumptions are you making? Which ones could be wrong?”
Self-reflection uses more tokens but catches mistakes before they reach code review. The net cost is lower.
Problem Decomposition
When one-shotting a complex task fails, break it down:
Big Task A
├── A1: Extract the data layer
├── A2: Write the transformation logic
├── A3: Add error handling
└── B: Wire up the API endpoint
CLAUDE.md as a Self-Updating Artifact
Every time you correct Claude, tell it:
“Update CLAUDE.md so you don’t make that mistake again.”
Claude is good at writing rules for itself. Over time, your CLAUDE.md accumulates project-specific knowledge that prevents repeated mistakes. This is the compounding advantage of agentic engineering — the context improves with every session.
Behavior Injection for Agents
Agent behavior affects output quality as much as the prompt content. A “thinks methodically” debugging agent produces different results than a “moves fast” prototyping agent.
Set behavioral constraints in agent definitions or CLAUDE.md sections:
| Context | Behavior to inject | Why |
|---|
| Debugging | ”Think step-by-step. Verify each hypothesis before moving on.” | Prevents shotgun fixes |
| Architecture | ”Consider three approaches before recommending one. Explain tradeoffs.” | Forces exploration |
| Code review | ”Be skeptical. Assume bugs exist. Check edge cases.” | Counteracts optimism bias |
| Prototyping | ”Favor speed over perfection. Use placeholders for non-critical details.” | Prevents over-engineering |
/voice to enable built-in push-to-talk, then hold Space to record. You can rebind to a modifier combo like meta+k to skip the warmup delay.
Don’t worry about perfect transcription. Claude handles mistranscriptions, filler words, and informal language well. The speed gain outweighs the occasional misparse.
← Prev: Parallel Execution
