@ohld
AI coding agent workflow in 2026 (for those still in denial)
1. You get an idea for something non-trivial → dump it into Claude Code and say:
Let’s have a discussion. Interview me.
2. Ask it to generate a proper spec https://github.com/github/spec-kit + implementation plan (Spec Kit–style) and save everything into specs/*.md:
Update specs/*.md and create an implementation plan with references.
3. Open the spec and plan manually. Tune it. Cut bullshit. Clarify intent.
4. Start a new chat. Clean context. No context rot. No compaction debt.
5. Create http://prompt.md and write:
Study specs/readme.md.
Study specs/my-new-feature.md. Pick the most important thing to do next.
6. Optionally add:
• write property-based or unit tests (whatever fits)
• run tests after changes
• prepare for deploy
7. Then run the loop:
while :; do cat http://prompt.md | claude --dangerously-skip-permissions; done
8. If the agent stalls – edit http://prompt.md or the spec.
Deal. share your successes!
0 reply
0 recast
0 reaction