When to use Fable 5.2 vs Opus in Claude Code, plus prompt-audit

Peter Yang's rule for choosing between Fable 5.2 and Opus in Claude Code is to keep Opus as the everyday model and save Fable for the hardest and most important work. He sets it out in a video tutorial that he says is sponsored by Anthropic, which makes both models, and he says he has spent more than 500 hours using them. He accepts that Fable scores better in almost every category of Anthropic's launch benchmarks, singling out agentic coding and long-running open-ended work, but says Opus is more than enough for most knowledge work, especially execution. Cost is his other reason. He says Max subscribers can spend up to half of their weekly limit on Fable, and that on the API Fable costs twice as much per token. His default matches Anthropic's own advice: its Fable 5.2 launch description recommends starting most coding and knowledge tasks on Opus 5.5 and calls Fable best for the most ambitious, open-ended technical work, and the company has published its own guide to choosing between the two. Yang stretches Fable further, to personal decisions where he wants the strongest model.

The video applies the rule to five of his own setups. A chief of staff that triages his email and calendar runs on Opus, because he says that job does not need the most powerful model. An advisor for career and life decisions runs on Fable, grounded in a personal plan document listing his goals, his principles and what gives or drains his energy. Automated workflows, such as the chain of skills that produces his weekly podcast, run on Opus unless they get very complicated. Personal apps are built with Opus, which he calls good at execution, but he switches to Fable to hunt for bugs and plan big features. Asked to find bugs in a fitness app he had been refining with Opus, Fable listed 30 issues sorted by priority, which he calls eye-opening. The last setup, coordinating his mother's cancer care and researching clinical trials, runs on Fable because, he says, nothing matters more to him. He also passes on four tips for Fable that he credits to Alex from Anthropic: ask it what is wrong with an idea before it builds anything, start on medium or high effort rather than extra high or max, say up front what needs your approval, and give it a whole project rather than a single next step. In earlier head-to-head tests of the two models across five tasks, he found Fable ahead on several, with Opus still winning one.

The second half of the title is a prompt-audit command, which Yang calls a secret that most people do not know about. Before, he would ask Claude after each conversation how a skill could be improved, but he says that overfits to one example. Instead he runs "/claude-api prompt-audit" (the automatic captions render it as "Claude-API space prompt audit") and has Claude look across every conversation that used his podcast and video skills for feedback he kept repeating. In his example, as far as the captions can be followed, it reported 149 sessions over the past 90 days touching those skills and proposed exact copy edits. He says not every suggestion is good, so he reads them all, accepts some and rejects others, because he does not want his skills to drift into text he no longer understands. He prefers to run the audit with Fable, or whatever model is newest.

His post on X adds a second command the video does not demonstrate: "/doctor prompt-audit," which he says audits a whole Claude Code setup, including the CLAUDE.md file, skills and commands, and proposes fixes, while the claude-api version audits one skill in depth. The attached screenshot is described as splitting a review of 65 user skills and 60 plugin skills across five parallel reviewers. He says he wishes Claude Code suggested running both every time Anthropic releases a new model. Neither the post nor the video explains where these commands come from beyond his description, and the assignments above are his working habits, not measured comparisons.