/a-self-learner

The same finding keeps coming back. Review after review, the same missing check, the same pattern you keep reintroducing, the same false positive you keep re-explaining. Nobody writes it down, so next month it is there again.

a-self-learner reads the accumulated review history and turns the repeats into proposed changes to your rules and your review skill, so the finding stops being found because it stops happening.

It works from .claude/reviews/review-issues.jsonl, the log your review skill has been writing, plus whatever feedback notes the project keeps. It groups by recurrence rather than by severity, because a low-severity thing caught eleven times is a missing rule, while a one-off critical bug is just a bug. Then it hands the actual edits to a-rules-optimizer or a-review-optimizer, whichever owns the file that needs to change.

It proposes. It never writes. Every change comes to you as a specific diff with the findings that justify it, and nothing lands until you approve that one. A skill that rewrote your own rules on its own evidence would be a skill you could not trust, and one bad inference would compound quietly across every future review.

The recurring false positives matter as much as the recurring bugs. A check that keeps firing on a decision you made deliberately is costing you attention every run, and the fix is to record the decision, not to keep dismissing it.

It needs history to work. On a fresh project there is nothing to learn from, so this is the last of the three to become useful, and it gets better the longer the log runs.

Ships in the dev-workflow-forge plugin, alongside a-review-optimizer and a-rules-optimizer.

/plugin marketplace add imisic/claude-marketplace
/plugin install dev-workflow-forge@imisic

Then the command is /a-self-learner.