Review

---
name: review
description: Reviews how Flow performed. Finds where rules failed, where friction repeated, and where the design has a gap. Records findings as study cases or workflow notes, globally.
---

# Flow review

!`flow audit index --quiet 2>&1 || true`

## Clear failure

The user rejected output, or a loaded rule did not fire.

1. Keep the offending output verbatim.
2. Name every rule that loaded and should have prevented the failure. `## Whether a rule loaded` says how to tell.
3. Record a study case: `flow cases issues`, then `flow cases new "<title>" --issue <issue>`. `~/.flow/references/study-cases.md` carries the body format.

## Suspected flaw

Friction hit twice, a rule fought the work, or a pattern looks wrong without a concrete failure.

1. Name the suspicion in one sentence.
2. Investigate: this session's conversation, `flow audit read` for a turn range from a past session, `flow audit ls` to list them. Past those, the raw transcripts at `~/.claude/projects/`.
3. Compare what happened against the rules that loaded, by `## Whether a rule loaded`. A rule that loaded and stayed silent is defective. A rule that never loaded is absent: the fault is whatever should have loaded it.
4. Record the finding in `~/.flow/workflow-notes.md`, dated, with the project.

## Whether a rule loaded

`grep '"loaded"' ~/.flow/logs/scorecards/<session-id>.jsonl` prints one line per `CLAUDE.md`, file it imports, or `.claude/rules/*.md` file that entered the session's context. A path-scoped rule loads only after a matching file is read, so its line can be missing. The current session's id is `$CLAUDE_CODE_SESSION_ID`, and `flow audit sessions` lists past ones. A skill leaves no line: find its invocation in the transcript.

## Recording

- **Study case** (`~/.flow/study-cases/`): a failure with an artifact. Written in 2 stages: the artifact and one line now, the analysis later.
- **Workflow note** (`~/.flow/workflow-notes.md`): friction, a gap, a pattern. One dated line.

Record, then return to the work.