A skill for AI coding agents that keeps you reading the code instead of waiting for it.
When an agent dumps ten files in one shot, you stop being a programmer and become a spectator. You lose the mental model of your own codebase, you can't debug it later, and you spend the gaps between prompts drifting. flowcode trades a slower start for a developer who understands every line and stays in flow.
- The agent writes one unit at a time, a function, a class, a small module, in the order a human would build it: types, then core logic, then side effects, then wiring.
- Each unit is proposed as a diff you accept, reject, or tweak.
- Hidden somewhere in each unit is a 4-character hex flag in a comment. You report it back. No flag, no next unit.
- Every hand-off comes with a real question about the approach, so you're reviewing, not just hunting for hex.
Flags are removed once a unit is accepted, so nothing lingers in your codebase.
Claude Code personal:
cp -r skills/flowcode ~/.claude/skills/flowcode
Per-project:
cp -r skills/flowcode .claude/skills/flowcode
Other agents (Codex, Cursor, etc.): point them at skills/flowcode/SKILL.md however they load instructions.
Works best with per-edit approval turned on. Auto-accept still works, but then the flag is the only thing standing between you and spectating.
MIT