Step 4 - Implement & Review

Now we execute the plan. This step is an iterative loop of Generate → Review → Refine.

The Execution Loop

For each item in your plan:

1. Select a single step
   Pick the next smallest meaningful unit of work.

2. Ask the AI to implement that step
   In your IDE, provide the step description and relevant files. Let the AI propose a diff.

3. Review the diff carefully
   Before accepting anything, check:

   • Does this actually implement the step?
   • Does it align with the spec and rules/config?
   • Did it touch files or logic it shouldn’t?

4. Apply and validate
   Once the diff looks good, apply it and run quick feedback loops:

   • Typecheck/compile.
   • Targeted tests.
   • Local manual checks where appropriate.

Micro-Correction vs. Perfection

AI suggestions are often 80–90% correct.

• Bad pattern: Repeatedly regenerating large blocks of code trying to get a "perfect" answer.
• Velocity pattern: Accept the mostly-correct version, then manually adjust the last 10–20%.

Your time is better spent editing than re-prompting endlessly.

Controlling Code Bloat

Left unchecked, AI tends to:

• Add redundant helpers.
• Duplicate logic instead of reusing existing utilities.
• Over-comment trivial code.

As the human in the loop, you should:

• Ask the AI to use existing abstractions where possible.
• Delete unnecessary code and comments.
• Simplify implementations that feel over-engineered.

When to Stop and Re-Spec

If you find yourself fighting the AI repeatedly:

• Re-read the spec—did something change or was something missing?
• Refine the plan to make steps smaller or clearer.
• Consider updating project rules to steer future generations.

It is often cheaper to adjust the spec or plan than to wrestle with code that doesn’t fit.