If you've ever typed "write me a REST API" into ChatGPT and got back something you had to completely rewrite — this post is for you. The problem isn't the AI. It's the prompt. After testing 200+ prompts across Claude, ChatGPT, and Gemini, I found one framework that consistently produces production-ready code: CRTSE. Every great AI prompt has 5 parts: Here are 5 prompts built with this framework that I use every week: The problem it solves: You need a complete, production-ready API endpoint but don't want to write all the boilerplate from scratch. Copy this prompt: You are a senior Node.js backend engineer with 10+ years of experience. Create a complete POST /api/users endpoint in TypeScript using Express + Zod validation + Prisma ORM + JWT auth middleware. Return full code for routes/users.ts and middleware/auth.ts. Handle duplicate email (409), invalid input (400), and server errors (500). Include JSDoc comments. No external dependencies beyond Express, Zod, Prisma, and jsonwebtoken. Why it works: The role assignment ("senior Node.js backend engineer") activates the model's most relevant knowledge. The standards section ("no external dependencies") prevents bloated output. The error handling specification forces production-level thinking. Result: Full working code in one shot. No rewriting needed. The problem it solves: Your Node.js service memory grows over time and you can't find why. Copy this prompt: You are a Node.js memory profiling expert specializing in V8 heap analysis and garbage
collection. Review this code: [paste your code].
Identify: 1) All memory leak sources with exact
line references, 2) Why V8 GC cannot collect them, 3) Fixed version of each leaky function with before/after comparison, 4) Exact commands to profile with --inspect and take heap snapshots, 5) How to set up memory monitoring alerts in production. Why it works: Asking for "exact line references" forces precision. Asking for "before/after comparison" ensures you get actionable fixes, not just analysis. Result: I used this last week and found 3 event listener leaks in 2 minutes that I had been chasing for hours. The problem it solves: Your database queries are running slowly in production and you don't know why. Copy this prompt: You are a PostgreSQL performance engineer. Here is my slow query and EXPLAIN ANALYZE
output: [paste query + EXPLAIN output].
Provide: 1) Identified bottlenecks (seq scans,
missing indexes, bad joins), 2) CREATE INDEX statements to add, 3) Optimized query rewrite, 4) Expected performance improvement percentage, 5) Any schema changes recommended. Target sub-100ms execution time. Why it works: Giving the AI the EXPLAIN ANALYZE output is the key. It stops the AI from guessing and forces it to analyze your actual query plan. Result: Cut a 4-second query down to 60ms using the indexes this prompt recommended. The problem it solves: You want a security review of your backend before pushing to production. Copy this prompt: You are an application security engineer with OWASP expertise. Security audit this code: [paste code]. For each vulnerability: 1) OWASP Top 10 category, 2) Severity rating (Critical/High/Medium/Low), 3) Exact vulnerable line with explanation, 4) Fixed code snippet. End with a security hardening checklist. Priority order: SQL injection, auth bypass, sensitive data exposure, broken access control. Why it works: The OWASP category requirement forces structured output. The severity rating tells you what to fix first. The priority order at the end means critical issues always surface first. Result: Found a JWT verification bypass in my code that would have been a critical production vulnerability. The problem it solves: You inherited legacy code and need a clear refactoring plan without breaking everything. Copy this prompt: You are a software architect who applies SOLID principles and clean code practices. Analyze
this code for code smells: [paste code]. For
each smell: 1) Formal name (e.g. Long Method,
God Class), 2) SOLID principle violated, 3) Refactored version using the appropriate design pattern, 4) Risk level of refactoring (High/Medium/Low). Prioritize smells that will cause the most maintenance pain in 6 months. Why it works: Asking for "formal names" gives you vocabulary to discuss with your team. The "6 months" framing forces the AI to think about long-term maintainability, not just cosmetic fixes. Result: Turned a 400-line God Class into 4 focused services with clear responsibilities. Look at what every prompt above has in common: ✅ Specific expert role — not just "you are an expert" but "you are a PostgreSQL performance engineer" ✅ Numbered output format — forces structured, scannable responses ✅ Quality constraints — sub-100ms target, OWASP categories, severity ratings ✅ Edge case handling — duplicate emails, server errors, null inputs ✅ Actionable deliverable — not analysis, but working code or specific commands This is the CRTSE framework and it works across Claude, ChatGPT, Gemini, and GitHub Copilot. I packaged 50 of these prompts into a PDF covering 5 categories:
→ Code Generation (10 prompts)
→ Debugging (10 prompts)
→ Code Review (10 prompts)
→ Documentation (10 prompts)
→ Refactoring (10 prompts)
Each prompt follows the full CRTSE framework with a complete copy-paste ready example. Get the full pack here → Ai Prompts for Developers What's the best AI prompt you've used for coding? Drop it in the comments — I read every one. 👇