What it does
A meta-skill for when the system makes a mistake. You describe what went wrong, and Claude traces the root cause — was a rule missing? In the wrong place? Based on bad data? Then it patches the relevant skill instructions, CLAUDE.md, or memory files so the same mistake can't happen again.
This is how Patina gets better over time. Every error is a chance to tighten the instructions.
Connections used
None. Reflect reads and modifies skill instructions and configuration files.
How to use
Claude will ask what went wrong, then investigate and fix it.
How it works
- Identify the mistake in one sentence (no hedging)
- Trace the root cause: rule missing, rule in wrong place, bad inference, or data issue
- Classify where the fix belongs:
- Skill instructions (strongest — loaded during the workflow)
- CLAUDE.md (global rules)
- Memory files (context for future sessions)
- Implement the fix in the right file and the right step of the workflow
- Summarize: what happened, why, what changed, and why it fixes it
Design philosophy
Skill instructions are prioritized over memory because they're loaded during the workflow and can't be missed. Rules are placed at the exact step where they're needed, not in a preamble that gets forgotten. If a pattern repeats despite a fix, it escalates to a stronger mechanism.