Skip to content

Workflow: AI-Assisted Coding Session

A step-by-step workflow for a single AI-assisted coding session, incorporating all eight patterns that separate productive development from expensive flailing.

Based on Chapter 5: Building with AI

When to Use This Workflow

Use this every time you write or modify code with AI assistance -- Claude Code, GitHub Copilot, Cursor, or any AI coding tool. A randomized controlled trial found developers were 19% slower with AI assistance, but believed they were 20% faster. This workflow closes that perception gap.

The Workflow

Phase 1: Before You Start

Step 1: Set context (Pattern 1: Context First). Before your first prompt, tell the AI: your tech stack, the problem, constraints, and what you have tried. 150-250 tokens of context hits the sweet spot. If your project has a CLAUDE.md or rules file, confirm it is current.

Step 2: Checkpoint (Pattern 7: Checkpoint Commits). Commit your working code before any AI interaction: git add -A && git commit -m "checkpoint: before AI session". You will checkpoint again after each successful change.

Step 3: Plan the architecture (Pattern 4: Architecture Ownership). Decide component boundaries, data models, API contracts, and error handling before asking AI to implement. You own structure -- 80-90% of AI-generated code suffers from "avoidance of refactors" when the AI decides structure independently.

Phase 2: The Coding Loop

Step 4: Small requests (Pattern 3: Iterative Refinement). One function per request. One concern per request. For a complete feature: 5-10 iterations. The discipline feels slower. It is faster.

Step 5: Show examples (Pattern 2: Concrete Examples). Paste an existing component that follows your conventions. "Function names use camelCase, constants use SCREAMING_SNAKE_CASE" beats "use consistent naming."

Step 6: Describe behavior (Pattern 5: Test-Driven Prompting). Say "it should find items instantly even with 10,000 entries" instead of "use a hash map with O(1) lookup." Let the AI choose the how.

Step 7: Review and checkpoint (Patterns 7+8: Checkpoint + Review Ruthlessly). Review like you are reviewing a junior developer. Check correctness, security, edge cases. Run tests. Only 55% of AI-generated code is secure. If it passes, commit immediately.

Decision point: Change working? Yes -- commit, return to Step 4. No -- continue to Step 8.

Phase 3: When Things Break

Step 8: Escalate errors (Pattern 6: Error Escalation). Share the error message with the AI. Let it propose a fix -- AI excels at debugging its own mistakes. If the first fix fails, provide more context. After 2-3 failed attempts at the same problem, take over manually. The AI is stuck.

Decision point: Fixed within 2-3 attempts? Yes -- review, checkpoint, return to Step 4. No -- fix manually, checkpoint, return to Step 4.

Phase 4: Close the Session

Step 9: Final review. Run the full test suite. Review the cumulative diff since your pre-session checkpoint. Check for hardcoded secrets and insecure defaults. Commit with a descriptive message.

Tips

  • Keep rules files actionable. "Write clean code" isn't a rule. "All public functions have JSDoc comments with param types" is.
  • Do not use an LLM for what a linter can do. Formatting and import ordering are tool problems.
  • Checkpoint commits prevent the biggest time waste -- redoing AI work after a bad change.
  • Start fresh after 30-50 messages. Context degrades in long sessions. Summarize progress and continue in a new conversation.