Skip to main content
Goal: Collaborate with the AI on design decisions before writing any code.

​
Moving to “Idea Space”

The critical instruction to use:
“Don’t make any code changes yet, just propose a plan”

​
Why This Matters

  • Prevents premature implementation
  • Keeps the AI in exploration mode
  • Allows you to course-correct before code is written
  • Reduces “helpful” refactors that break things

​
Operating at Every Level

The senior engineer must guide the AI across multiple architectural levels:

​
Level 1: Architecture

  • How does this fit into the overall system?
  • What modules/services are affected?
  • Are there cross-cutting concerns?

​
Level 2: Dependencies

  • What existing code do we need to modify?
  • What new dependencies might we need?
  • Are there circular dependency risks?

​
Level 3: Style Conventions

  • What patterns does this codebase follow?
  • Are there established conventions to match?
  • What’s the testing strategy?

​
Level 4: Naming Details

  • What naming conventions are used?
  • How should new functions/variables be named?
  • Does this match the existing vocabulary?

​
Connecting the Dots

Instead of taking a “leap of faith” with a confident-sounding AI, ground suggestions in your own knowledge:

​
Ask for Alternatives

"What are 2-3 different approaches we could take here?
What are the trade-offs of each?"

​
Request Pros and Cons

"You suggested using Redis for caching.
What are the downsides? What alternatives exist?"

​
Challenge Assumptions

"You assumed we need real-time updates.
Can we use polling instead? What would that look like?"
Key insight: The AI doesn’t know your constraints—you must surface them through questions.

​
Creating Spec Documents

Know when to move a long chat-based plan into a durable spec.md file for larger features.

​
When to Create a Spec

  • Feature will take >1 day to implement
  • Multiple engineers need to understand the plan
  • The design needs stakeholder review
  • You’ll need to reference it across sessions

​
Spec Template

# Feature: [Name]

## Problem
What problem does this solve?

## Solution
High-level approach

## Architecture
How does this fit into the system?

## Implementation Plan
1. Step 1
2. Step 2
3. Step 3

## Edge Cases
What could go wrong?

## Testing Strategy
How will we verify correctness?
Pro tip: Ask the AI to generate the spec from your conversation, then review and refine it.
Key Principle: Design in “idea space” with full flexibility. Lock it in before moving to “code space” where changes are expensive.