Kingstinct AI Playbook

[robertherber]

Core principles
I think with a set of core principles we can achieve as much as possible, without wasting time switching between tools etc.

• In all repos hav AI instructions available for various AIs, symlink them so we only need to update once:
  • .github/copilot-instructions.md
  • CLAUDE.md
  • .cursorrules (I know this is legacy for cursor - but I think it is most compatible with other AIs)
  • AGENTS.md
• Spend some time every day exploring scenarios with AI (consistency + not putting entire days into just exploring is key)
• Always improve the AI instructions. If something requires a follow-up prompt - get that into the AI instructions so we don't need to repeat that follow-up prompt again.
• Always try to provide the AI with guardrail instructions. If it has tests, linting and typechecks to iterate on before coming back with a
• Stay aware of which agentic level you're working with:
  • Using Github Copilot Agents or similar (letting it do a PR from start to finish) frees up our time drastically, and can let us stay on a way higher abstraction level.
  • Claude Code and in parts Copilot can work on something for a few minutes but needs input from time to time. We still need to be on code level, but can multitask a bit.
  • Code completion and smaller chat tasks is more like having a pair programmer, lets us speed up our workflow but we're at the code level and can probably just work on one task at a time.