Basics of Prompting Coding AI
AI coding tools' output quality changes greatly with how you ask. Here are coding-specific tips distinct from general prompts.
5 Golden Rules
1. Give Enough Context
Over "fix this code," say "this project uses Next.js 16, TypeScript strict, Vitest. Fix module X like Y."
2. State the Completion Condition
Over "make it nice," say "all tests pass, no lint errors, existing API compatible."
3. Convey Constraints
- "Libraries you may use: existing dependencies only"
- "Don't call external APIs"
- "O(N) or less"
- "Within 100 lines"
4. Use File References
- Cursor: reference with @filename
- Claude Code: "refer to @auth.ts"
- Copilot Chat: auto-references the file open in the editor
5. Ask for a Plan First
"First a 5-step implementation plan. Start implementing after I OK it. Ask if questions arise midway"
Frequent Patterns
1. New Feature Addition
Feature: users can upload a profile image Requirements: - Format: JPG/PNG, under 5MB - Storage: S3 - DB: save url in the profile table - Use the existing auth middleware - Add tests too First make a plan, implement after my OK.
2. Bug Fix
Symptom: top page 404s after login Repro steps: 1. Access /login 2. Enter email/password 3. Click "Log in" → /home is 404 Expected: /home displays normally Recent change: feat/auth-refactor PR Investigate and propose cause and fix.
3. Refactor
Target: all of lib/auth/* Goal: migrate to JWT (currently session) Constraints: - Maintain existing API compatibility - All existing tests pass - Staged: first run JWT in parallel, remove session later Present plan and impact range first.
4. Add Tests
Target: src/utils/parseDate.ts Coverage: 0% → 90% Test framework: Vitest Cases: - Normal (multiple formats) - Abnormal (invalid input, null, undefined) - Edge (leap year, timezone) After implementing, confirm coverage and report.
Patterns to Avoid
- "Make it better": vague, AI interprets arbitrarily
- "In the latest style": criterion unclear
- "Cool": over-seasoned
- A huge "all": split it
Language-Specific Tips
TypeScript
- State "no any, strict mode"
- Specify "type-validate with Zod / Valibot"
- Expected type inference when using generics
Python
- "Type hints with typing, mypy strict"
- "Manage with pyproject.toml and uv"
- "With Pydantic v2"
Rust
- "No unsafe blocks"
- "Address all Clippy warnings"
- "State lifetimes explicitly"
Recovery on Failure
Tell It "It Doesn't Work"
"Test X fails with the earlier code. Error: [full error message] Analyze the cause and fix"
When the Direction Is Wrong
"The approach is wrong. I want to go in the ~ direction. Discard existing code and rebuild from a plan"
In-Session Tips
- If too long, a new session (prevent context contamination)
- Commit big results each time
- Periodically confirm "summarize so far in 5 lines"
Persist with Skill / Rules
Put instructions you write every time in CLAUDE.md / .cursor/rules / .github/copilot-instructions.md. AI auto-follows them next time.
Next Step
To the deep-dive articles "Mobile App Development with Claude Code" and "Claude Code Agent Features."
