Skip to main content
Goal: Build the AI’s understanding of your codebase before asking it to make changes.

​
The “Refresh” Prompt

Participants must learn to explicitly direct the AI to build its own context rather than relying on “magic” retrieval.

​
Example Prompt

Before we start, please:
1. Read the main entry point (src/index.ts)
2. Review the authentication module (src/auth/)
3. Check the database schema (prisma/schema.prisma)
4. Summarize the current architecture
Why this works: You’re giving the AI a clear roadmap instead of hoping it finds the right files.

​
Verifying the Summary

The engineer must review the AI’s summary of the code to ensure it isn’t:
  • ❌ Looking at deprecated files
  • ❌ Missing key dependencies
  • ❌ Misunderstanding the architecture

​
Red Flags to Watch For

  • AI mentions files that don’t exist
  • AI describes patterns that aren’t actually used
  • AI misses critical dependencies or integrations

​
Watching Tool Calls

Senior engineers should monitor the AI’s “work”—the files it reads and searches it runs—to spot when it’s on the wrong track early.

​
What to Monitor

  • Which files is it reading?
  • What search queries is it running?
  • Is it exploring the right parts of the codebase?
Pro tip: If the AI is reading irrelevant files, interrupt and redirect immediately.

​
Managing Context Rot

Techniques for preventing the AI’s reasoning from degrading as the context window fills up:

​
Strategy 1: Architecture Files

Maintain high-level architecture.md files that provide:
  • System overview
  • Key design decisions
  • Important constraints
  • Module relationships

​
Strategy 2: Break Into Smaller Steps

Instead of one massive task, break into digestible chunks:
  • ✅ “First, let’s update the auth middleware”
  • ✅ “Next, we’ll add the new endpoint”
  • ✅ “Finally, update the tests”
Why: Keeps the context focused and prevents drift.

​
Strategy 3: Periodic Resets

For long sessions, periodically:
  1. Summarize progress
  2. Start a fresh conversation
  3. Provide the summary as context
  4. Continue with clear focus
Key Principle: The AI’s context is YOUR responsibility. Manage it actively, not passively.