Chanjoong (CJ) Kim

Seoul

All projects

reflect. A self-⁠correction harness for Claude Code.

When
Event
Anthropic Built with Opus 4.7 Hackathon
Stack
TypeScript, Claude Code hooks
A list of reflections, one of them open: the pattern it found, the signal that caught it, and the adjustment for the next turn.
The reflect Viewer, with the example session it ships with.

Why

When a coding agent goes wrong, the same signs keep coming back: a revert, a deleted file, a "no" from me. I figured that if those signs were scored, and a correction went into the next turn whenever the score crossed a line, the agent could stop repeating the mistake.

How

reflect never blocks the work: it keeps score in the background, and once the line is crossed it asks a model why the work keeps being rejected, then steers the next turn.

Outcome

I built reflect for the hackathon, where I was one of 500 selected from 20,000+ applicants, and published it on npm under the MIT license.