I found these tips very valuable for coding with AI:

  1. Be intentional about your AGENTS.md: know what it contains and keep it short.
  2. Create a skill called myVoice. Paste anything that has your authentic voice in it. Update your AGENTS.md to summon that skill whenever AI or a skill needs to use your voice.
  3. Update your AGENTS.md so that whenever you give feedback on the output of a skill, it ends with a suggestion for updating the skill with your new feedback.
  4. If using several providers at once, make sure they use shared resources (AGENTS.md and .agents/skills/). Configure each provider to read them, since discovery differs.
  5. Provide agents with default tools for saving tokens: Poppler, ImageMagick, LibreOffice, Pandoc, qpdf, Tesseract, ExifTool, FFmpeg, yt-dlp, whisper.cpp, ripgrep, jq, uv, Corepack, CMake, CodeGraph, ngrok.
  6. Fix what caused a mistake, beyond correcting the code. Whenever the AI didn't give you the output you expected, ask what led to it "Why did you do it that way?" or "Why did you forget that?", then make sure that can't happen again.

Engineering used to be about planting seeds: bringing a function, a file, a feature, or a codebase to life.

With AI, our work is about gardening: making sure those seeds, now created by AI, grow well, pruning what no longer belongs, and removing obstacles along the way.