Peter Yang demonstrates five personal Claude workflows and the model choices he makes for each. Anthropic sponsors this video; the comparison reflects his experience, not an independent benchmark.

Assign the expensive model to harder work

  • Yang says he uses Opus 5.5 for daily execution and Fable 5.2 for strategic advice, complex planning, difficult bug searches and care coordination.
  • He cites Anthropic’s launch benchmarks and says Fable costs twice as much per API token in the comparison shown. Pricing and access limits are time-sensitive source claims.
  • His suggested practice: ask for weaknesses before implementation, set approval boundaries, and give the model a whole project when long-horizon work warrants it.

Five illustrated workflows

  1. Mail and calendar: a twice-daily chief of staff ranks incoming work; Yang reviews and edits drafted messages before sending.
  2. Career advisor: a human-authored plan with goals and sources of energy grounds blunt advice. The output is a prompt example, not a verified judgment.
  3. Podcast production: separate skills prepare interview guides, edit instructions, short videos and newsletters; Yang still chooses clips, music and final edits.
  4. Workout app: he iterates on the app for about a month, uses a visual canvas for designs, and asks the stronger model to suggest bugs. The 30 reported issues are suggestions, not 30 verified defects.
  5. Care brief: an example one-page document organizes contacts, questions and clinical-trial leads for a doctor visit. The medical information shown is fictional; clinicians must verify actual eligibility and treatment decisions.

Improve skills from repeated feedback

  • Yang asks the assistant to review many past conversations for recurring corrections and propose exact before/after skill edits.
  • He signs off on selected changes instead of letting one odd conversation permanently rewrite the workflow.
  • Caveat: no controlled comparison tests Fable against Opus on these five tasks; the strongest medical and productivity claims remain creator-presented demonstrations.

“I usually don’t get AI to just send the email by itself because I want to review it.” — Peter Yang, 5:11