Quick one-liner suggestions — overkill for small tasks
MCPFull-repo
The 8 Core Lessons
1
POC ≠ foundation
Use AI Studio / Claude Artifacts to prove the idea fast. Then throw the code away. Extract a spec, start fresh. Never refactor a POC into production code.
2
AI knowledge has an expiration date
For any external API, feed the current docs. AI training data is frozen — it will confidently generate deprecated code. Always verify against the source.
3
Planning is the feature
Always produce a spec or design doc before asking for code. Ask for a plan first, review it, then implement. The planning step costs 5 minutes and saves hours of debugging.
4
Context is the prompt
AI doesn't see your system unless you show it. Paste your data shapes, existing module, and constraints every time. Anchoring is the skill that separates useful output from hallucinated output.
5
Scope or suffer
"Make it better" is not a spec. Name the files to touch, the files to leave alone, and the behavior that must not change. Commit guardrails to CLAUDE.md or .cursor/rules so they apply to every session.
6
Decompose before you delegate
Handing AI a large task produces large tangled output. Break work into atomic units — one function, one file, one concern — before prompting. Small inputs produce reviewable, mergeable outputs.
7
Tests are your receipt
AI-generated code without tests is a promise with no proof. Write or generate tests alongside implementation, not after. Tests also make refactoring safe — ask AI to write them before it changes anything.
8
Leave a paper trail
AI has no memory between sessions. CLAUDE.md, .cursor/rules, decision logs, and artifact files are its persistent memory. If it's not written down, the next session starts from zero. Your files are the brain.