Fable review skill replaces CodeRabbit re-reviews

The setup grew out of a specific mismatch in the engineer's tool mix. Claude on the Max plan had enough headroom to add more review agents, while CodeRabbit on the Team plan kept re-hitting rate limits — days when the interface still showed '2 times per hour' and waits of 40–50 minutes even after the upgrade. With six repositories for two products, each app fix spawning a paired E2E test PR plus documentation PRs, the PR count was high relative to the amount of work.

The engineer had already cut automatic reviews so they only run on request and had one E2E repo reviewed by Fable instead of CodeRabbit. The remaining waste was two cases: re-reviewing a one-line fix and sending docs-only PRs through CodeRabbit. The first use of the new skill on a docs-only PR returned 'fix first' with four findings, each a mismatch inside the documents, such as an outdated skill count and a missing update-log row. After the fixes, a second look at only the increment returned 'pass' — showing the skill catches consistency issues that a quick visual scan of a diff tends to miss.

In the design, the engineer drew the line first: code-containing PRs must pass CodeRabbit at least once, and only the later increments and docs are eligible for Fable. The guardrail sits in the scripts, not just the prompt: unknown verdict text or a missing marker becomes a fail, and merge checks only accept a verdict matching the current 40-digits head. The engineer's own takeaway is that combining AI tools makes the most rate-limited tool set the pace for the whole process.