Skip to main content
Time Investment: ~5% of total workflow Once the Rehearsal reaches 90-95% confidence, the engineer’s role shifts from collaborator to active observer.

​
The Switch to Observer

Your role during execution:
  • ✅ Monitor progress and tool calls
  • ✅ Watch for signs of trouble early
  • ✅ Interrupt if the AI goes off-track
  • ❌ Not hands-off autonomy—active observation
Mindset shift: From “let me think with you” to “let me watch you work.”

​
Direct Implementation

The implementation prompt should be simple and direct.

​
Example Prompt

Based on our plan, please implement this feature cleanly,
concisely, and elegantly. Follow the architecture we discussed.

​
What “Clean, Concise, Elegant” Means

  • Clean: No dead code, no debug statements
  • Concise: Minimal complexity, no over-engineering
  • Elegant: Readable, maintainable, idiomatic
Pro tip: Reference the confidence-tuned plan explicitly to keep the AI grounded.

​
Active Observation

Watch for signs of trouble:

​
Red Flag #1: Repetitive Behavior

The AI keeps trying the same edit multiple times What it means: Context window is full or it’s confused What to do: Stop and refresh context

​
Red Flag #2: File Edit Struggles

The AI can’t successfully edit a specific file What it means: File too large, syntax issues, or model confusion What to do: Break into smaller edits or provide file structure

​
Red Flag #3: Off-Plan Implementation

The AI is making changes not in the rehearsed plan What it means: Lost track of the plan or misunderstood What to do: Interrupt and re-anchor to the plan

​
Red Flag #4: Stale Information

The AI references old file contents or deprecated patterns What it means: Context hasn’t been refreshed What to do: Re-run bootstrapping to rebuild context

​
Interrupting for Quality

Critical rule: If implementation goes off the rails, stop immediately.

​
When to Interrupt

  • Context window is full (model becoming incoherent)
  • Stale information being used
  • Multiple failed edit attempts
  • Deviating from the rehearsed plan

​
How to Interrupt

Stop. Let's pause here.
[Explain what went wrong]
[Provide corrective guidance]
Let's refresh context and try again.
Don’t let it spiral: Five minutes of bad output requires an hour of cleanup.

​
A Word of Caution: Shell Commands

⚠️ CRITICAL SAFETY RULE
  • ✅ Code changes can be rolled back via git
  • ❌ Shell/terminal commands cannot be undone
  • 🚨 Terminal commands require hyper-vigilance

​
High-Risk Commands

  • rm -rf (deletion)
  • git push --force (history rewriting)
  • Database migrations (schema changes)
  • Deployment scripts (production changes)
  • npm publish (package publication)

​
Safety Protocol for Shell Commands

  1. Read the command before execution
  2. Understand what it does
  3. Verify it matches the plan
  4. Check for destructive operations
  5. Only then approve
Best practice: Ask the AI to explain risky commands before running them.
Key Principle: Implementation is a “coherent dream”—watch the first few seconds carefully. If it starts wrong, it won’t self-correct.