| name | elegance-review |
| description | Run the BLIND frame-breaking elegance review on a plan or implementation via the elegance-reviewer agent. Use AFTER a correctness review (plan-review-invoke, plan-review-fanout, or codex-plan-review) and BEFORE execution, on any plan that creates new files, scripts, abstractions, or pipelines. Accepts a path (plan file, script, or directory) as the argument. NOT for: style trimming of fresh diffs (pr-review-toolkit:code-simplifier), correctness review (plan-reviewer), or reviews where the author's rationale must be weighed (this skill deliberately withholds it).
|
| argument-hint | <path-to-plan-or-code> |
elegance-review
Invokes the elegance-reviewer agent (.claude/agents/elegance-reviewer.md) correctly.
The one rule that makes it work: the agent is invoked BLIND.
Self-preference research says a reviewer that can recognize the author's reasoning rates the work higher, so the isolation must be structural.
Steps
- Resolve
$ARGUMENTS to the artifact path. If empty, ask for it; never guess.
- Build the agent prompt with ONLY:
- the artifact path(s) to read;
- the hard constraints the solution must satisfy, stated as bare facts;
- the repo's reuse surfaces worth searching (existing scripts, helpers, tests).
NEVER include: why the approach was chosen, the planning conversation, prior review verdicts, or your own opinion of the artifact.
- Spawn the
elegance-reviewer agent with that prompt. Run it in the background if a correctness review is running in parallel; the two must not see each other's output before both finish.
- When the report returns: archive it under
.claude/codex-reviews/YYYY-MM-DD-elegance-<slug>.md (same shelf as the other review transcripts), then present the ranked counter-proposals to the user for adopt/reject rulings. Adopted counters are recorded where the artifact's decisions live (the repository's documented decision location, or the reviewed artifact itself), never silently applied.
Design rationale
.claude/reference/elegance-reviewer-design.md maps every element of the agent to its evidence.
Division of labor across the review tools is recorded there too.