AI·

AI Pairing: Notes to Self

My ever-evolving tips for pairing with AI - not rules, just habits I keep forgetting
A software developer at a dual-monitor workstation, one screen showing code while the other displays an ethereal, translucent AI assistant figure made of flowing light particles and neural network patterns, both entities focused on the same problem. Warm desk lamp casting golden light from the left contrasts with cool blue-white emanating from the AI presence, creating dramatic interplay of shadows across scattered sticky notes and a well-worn notebook. Cinematic realism with subtle sci-fi elements, 8k photorealistic quality, shallow depth of field focusing on the collaborative space between human and AI, teal and amber color grading, overhead three-quarter angle composition.

This is a living document. Not "best practices" - just things I keep re-learning. The landscape changes weekly.

The Mindset Shifts

  • You're the architect, AI is the builder - I still catch myself letting AI drive when I should be steering
  • Better context = better output - Every time I think "that's obvious", I'm wrong. Say it anyway.
  • Talk like you're pairing with a senior dev who just joined- They're smart but don't know your codebase
    • use Claude.md, craft clean structure and conventions and it'll pick them up fast
  • Plan Mode for your problems, Edit Mode for Claude's

    "Plan mode for your problems, Edit mode for Claude's problems."

Habits I Keep Forgetting

  • Use voice/dictation more - typing less context because lazy
  • Screenshot the error, don't describe it
  • Create an issue when distracted by improvement ideas - don't derail current task
  • ctrl+z to terminal, fg to resume - way better than ! for throwaway commands

What's Working Right Now

  • Plan mode before any non-trivial change
    • easier to course-correct early than after tons of code is generated
  • Subagents for everything
    • Keith pushed me to use them for exploration, but now I use them for everything - main agent acts as orchestrator
    • I've had a lot of luck with this - larger tasks can run autonomously for a long time
  • Web search to load current docs into context
  • Read every line of changes before committing - easy when new, gets sloppy fast
    • AI adds the worst type of errors - small, subtle, hard to catch ones

What I'm Experimenting With

  • More aggressive use of CLAUDE.md for project conventions
  • Custom slash commands for repeated workflows
  • Letting AI write the first draft of issues/PRs

Lessons Learned the Hard Way

what is this

  • Don't trust AI-generated paths without verification
  • If stuck in a loop, start fresh context rather than fighting
  • Once you get "You're absolutely right" from AI, you're probably off-track - start a new session

Last updated: 2025-12-16