AI as a tool.
Not a crutch.
Use AI to ship faster without outsourcing understanding. One rule: if you can't explain every line, you don't ship it.
The rule: explain every line
If you can't explain every line of AI-generated code, you don't ship it. You own the code, not the AI.
"Treat AI-generated code as a draft that requires human review and understanding." — Addy Osmani
Measured risk
PRs containing AI-authored code had 1.7x more issues than human-only PRs. Review isn't optional.
What you gain
Accountability, preserved skills, fewer hidden defects, and less team friction when you own every line.
When you skip review
Lost accountability, eroded skills, hidden defects that cost 30–100× more to fix in production, and reviewers catching issues you should have found. Your job shifts from writing code to integrating it.
Four AI workflows (tRPC project)
Use these four workflows on your T3 project. Each has a clear input and output.
Comprehension
Paste unfamiliar tRPC middleware. Ask: "Explain this in plain English". Use to onboard to existing patterns.
Debugging
Structured prompt: error message + what you tried + context. Get targeted help instead of generic fixes.
Boilerplate
"Generate a Zod schema for a user profile form with these fields" → review and adjust. Don't hand-edit blindly.
Design rubber-ducking
"I need tagging on tasks. Separate table or JSON field? Here's my schema." Discuss tradeoffs before coding.
Live demo
Run through all four workflows on your T3 project during the session.
Rubber-duck rule
Design prompts are for thinking out loud. Don't implement until you understand the tradeoffs.
The prompting formula
Context + Constraint + Specific Ask. 61% of developers struggle with AI code quality due to insufficient context—not model limits.
Bad
"fix my code"
No context. No error. No behavior. AI guesses.
Good
"Here's my tRPC router [paste]. The create mutation throws [error] when I send [input]. I expect [behavior]. What's wrong?"
Context + constraint + specific ask.
Research insight
Context improves code generation more than role-playing personas. Treat AI like a colleague who needs onboarding to your codebase.
When not to use AI
AI is a tool. Some situations require human judgment and full understanding.
Architectural decisions
Don't use AI for decisions you don't understand yet. Learn first, then decide.
Security-sensitive code
Auth, tokens, encryption. If it guards user data, you must own every line.
Tempted to skip understanding
If your instinct is to paste and hope, stop. That's when you need to slow down.
Never outsource the reading. Hold AI-written code to the same standards as human teammates.
Claude Maxxing
Skills, MCPs, context engineering, and autonomous loops. Get the most out of Claude Code.
Explore
claudemaxxing.org — Skills, MCPs, CoVe (Chain-of-Verification), Ralph Loop, GSD (Get Shit Done).
Chain-of-Verification
Create → Outline → Verify → Emit. Claude tests its own code against docs and codebase.
Get Shit Done
Discuss, plan, execute, verify. State in files, not context. Fresh 200K windows per task.
"Give me six hours to implement a feature and I will spend the first four writing the prompt." — adapted from Eisenhower