Setup:
- For instruction files I use github copilot instruction file system since it does directory level instructions. So I can have instructions that only get added to the context when the agent is modifying code in a particular directory. My CLAUDE.md and AGENTS.md files just point to the github copilot instruction system and how to use it: https://docs.github.com/en/copilot/concepts/prompting-and-co...
- For MCPs the only two I use are cclsp (LSP MCP that works well on all of them) and playwright.
- For skills I have used skills.sh a lot for finding new skills, and the remotion skills have helped a lot with video editing (I was using davinci resolve for a while, but remotion lets the agent just do it and omg it's nice)
- I think I'm too trusting but I always do the --dangerously-skip-permissions option for permissions (yolo). I always use them in a repo, so I can revert any changes, and I usually turn the sandbox feature on or just checkout my repo in a docker container (my way to manually sandbox that works for every harness).
- I've built a tool that remembers the harness specific flags and sends the prompt for me, so I can send a prompt to codex, claude, or any of others using a single interface.
- I Have NOT tried cursor yet.
Oh, question answers:
1) I pay the $20 claude and codex, and the $10 copilot, so $50 total a month on AI.
2) For qwen, opencode, and agy I do the pay as you go (since qwen canceled there awesome free plan :( )
3) CLIs all the way!! (mainly because the tool I built to send prompts to them only works with the CLI)
4) Something custom, I can share if you like, it's still a work in progress, maybe I'll post to Show sometime, I'm still new here
5) Claude's cowork thingy and claude's app and web interface but no scheduling or always on thingy yet