Turn an agent's session mistakes into a rules skill for another model

Botirjon Mahamadaliyev spent a few days on a Pro plan working with Claude Fable 5.2. Together they built and published a Dart payments package on pub.dev and two agent skills, with READMEs, CI, marketplace listings and articles. Before switching back to Opus 5.5, he asked Fable to write a skill capturing its working habits. Each of the 17 rules is backed by a receipt: something that actually went wrong in their session and how it was caught. Examples include a pub.dev setting that never saved because a confirm dialog was skipped, invisible Unicode characters found when the skills linter ran on its own source, and screenshots that cost about 2.7 million tokens by staying in context. The skill was published to GitHub as fable-receipts.

Three projects were published in a few days. Fable hit the Pro plan's five-hour limit in the middle of a publishing task. On the first day with the skill, Opus 5.5 created the GitHub repository and dropped a few characters from the description; the read-it-back rule caught it, and that moment became the demo on the marketplace listing. The author has not benchmarked Opus with and without the skill.

The author says plainly that a skill changes habits, not capability. The harness was not named in the post, though one rule is about reading Claude Code logs. Deep reasoning on a Pro plan burns the usage limit fast, so plan for the session to stop mid-task.