Peter Yang's full walkthrough answers a practical question about Claude's two most capable models: which one should do the daily work, and when is the more expensive model worth calling in? He shows five workflows from his own Claude Code desktop setup: an inbox chief of staff, a goals-based advisor, podcast production, a personal workout app, and a care brief for a loved one. He says he has spent more than 500 hours with the models. The video is sponsored by Anthropic, so his demonstrations are useful examples, not an independent comparison study.
Use Opus 5.5 for routine execution and iterative builds. Escalate to Fable 5.2 for ambiguous, high-effort planning, a difficult audit, or a consequential question that deserves a stronger second pass. This is Peter's workflow, not a claim that either model is infallible. Anthropic also recommends starting with Opus for most coding and knowledge tasks.
Watch the Full Guide
Start With a Small, Reviewable Workspace
Peter works mostly in the Code tab of the Claude desktop app, keeping separate threads for his chief of staff, advisor, podcast and app projects. In Customize, he connects the tools each thread actually needs: Google Workspace for the inbox/calendar job; Riverside, Suno and his own files for production. Claude Code's official page emphasizes that the user sets permissions and reviews work. A connector can expose sensitive mail, meetings or health documents, so connect only the accounts needed for a given workflow and inspect permission scopes before enabling them.
| Job | Peter's default | Why | Human checkpoint |
|---|---|---|---|
| Inbox and calendar | Opus 5.5 | Structured daily triage | Approve every send or calendar change |
| Career and life advisor | Fable 5.2 | Ambiguous, strategic judgment | Check inferences against your own goals and evidence |
| Podcast production | Opus 5.5 | Repeatable steps and revisions | Review clips, music rights and final edit |
| Personal mobile app | Opus 5.5 to build; Fable for harder review | Iterate cheaply, escalate difficult bugs | Test screens and verify each reported issue |
| Care coordination | Fable 5.2 | Complex document synthesis and question preparation | Patient consent and clinician review for medical decisions |
At the API list price checked on 12 November 2026, Opus is $4 input / $20 output per million tokens, while Fable is $8 / $40. Subscription limits are a separate product rule and can change; do not treat API prices as a direct measure of what one desktop run cost Peter.
1. An Inbox and Calendar Chief of Staff (Opus)
Peter's chief of staff reads recent mail, sent mail and the calendar, then ranks open loops. He replies to the numbered list with instructions such as "draft a response" or "research this guest." He edits the drafts before any email goes out. After validating the briefing, he schedules it for morning and late afternoon; the automation surfaces work, but should not silently send messages or move meetings.
Peter's prompt:
Be my chief of staff. Check my email, calendar, and sent mail from the last 7 days. Find anything I owe someone, anything someone owes me, and any meeting I need to prep for in the next two days. Give me a numbered list, most urgent first, with a link to each email. Don't send or change anything without asking me first.
For a first run, use a test or low-stakes account if possible. Verify that each item links to the right thread, that dates and obligations are correct, and that stale messages are not presented as current tasks. Only then consider a scheduled briefing.
2. A Career and Life Advisor (Fable)
The advisor is grounded in a personal plan rather than a chat history alone. Peter documents his goals, principles, and what gives or drains energy, then asks Fable to compare current activity with that plan. In the demo, it challenges the gap between his subscriber goal and actual pace, and questions whether his product assumes customers enjoy tinkering as much as he does. This is a prompt for reflection, not proof that an AI has discovered a hidden truth about your character.
Peter's prompt:
Be my advisor. How am I doing relative to my goals, and what should I focus on? Also tell me something special or unique you've noticed about me that you think I haven't realized about myself yet. It doesn't have to be positive, and you don't have to be nice. Just be truthful.
A practical improvement is to ask it to cite the plan or dated activity behind each conclusion and to mark unsupported impressions as hypotheses. Keep the personal plan current, and do not outsource major life decisions to a single generated answer.
3. Produce a Podcast Without Losing Editorial Control (Opus)
Peter breaks his weekly episode into skills for interview preparation, selecting highlights, editing notes, show notes, newsletter and social clips. In the video, a Riverside connector gives Claude access to a recorded interview. Opus selects moments for an intro reel; Peter then chooses a Suno music option, supplies B-roll references and reviews the cut. He built a small personal video review app so he can attach comments to the exact frame instead of writing a long list of timestamps.
The lesson is the workflow architecture, not the claim that one prompt replaces an editor. Keep a source log for footage and music, check licenses and participant consent, and watch the exported cut for context, captions and audio. Peter says a particular reel took about 20 minutes of his interaction time; that is his example, not a general production estimate.
4. Design and Build a Personal Mobile App (Opus, Then Fable)
Peter's workout tracker began with his own need and took roughly a month of use and iteration, not a single prompt. He uses Claude Design's canvas to explore screens and revise copy, then asks Opus to implement the chosen direction. For a deeper review he switches to Fable, asking it to find bugs and UI cleanup opportunities. In the demo Fable returns a long prioritized list, but a reported issue still needs reproduction and a test before it becomes a fix.
A reproducible version of the loop is: define one user journey, generate two design options, select one, build the smallest working screen, test it on a phone, log failures, and ask Fable to audit only the difficult or unresolved parts. Keep a commit before broad changes so the review can be checked against actual behavior.
5. Organize Care for a Loved One (Fable, With a Medical Team)
Peter shares a sensitive example from helping his mother through cancer care. His open-source Fuck Cancer skill builds a one-page brief with contacts, next actions, confirmed facts, questions, medical terms and a care log. The video uses sample patient information rather than his mother's real medical record. The repository can save a local Markdown file or, with a connected account, a shareable Google Doc. Review its files and permissions before installation, and ask the patient what may be stored or shared.
The skill can gather possible trial questions from official records, but a search result is not an eligibility decision or treatment recommendation. Peter's README says the skill does not diagnose, choose treatment, determine trial eligibility or replace clinicians. The National Cancer Institute's trial-search guide recommends starting with the care team and comparing diagnosis and history with each trial's criteria; ClinicalTrials.gov notes that trials have distinct eligibility rules and may not benefit a participant. Use the brief to prepare better questions for the treating clinician, with source links and dates attached.
The Extra Workflow: Improve Skills Across Many Sessions
At the end, Peter audits repeated feedback across podcast and video conversations rather than changing a skill after one imperfect run. He asks for exact proposed edits, rejects the ones that do not fit, and only then updates the skill. The command shown is /claude-api prompt-audit; treat that as the command in Peter's setup, not a guaranteed built-in command in every Claude installation. If your environment does not expose it, ask Claude to review the relevant conversations or exported feedback and propose a diff without applying it.
Peter's prompt:
/claude-api prompt-audit Review all the conversations that we've had about the podcast and about the video skills, and then give me an updated list of changes that I should make to improve my skills based on my feedback in all these threads.
Keep the audit bounded to the conversations you intend to share. Ask it to group recurring feedback, cite examples, show exact wording changes and leave the skill untouched until you approve them.
Video Chapters
| Time | Segment |
|---|---|
| 00:00 | Fable 5.2 vs Opus 5.5 |
| 02:32 | Claude Code desktop setup |
| 03:30 | Inbox and calendar chief of staff |
| 06:11 | Career and life advisor |
| 09:14 | End-to-end podcast workflow |
| 14:01 | Personal mobile app and Claude Design |
| 18:47 | Care coordination for a loved one |
| 20:54 | Prompt audit for improving skills |
Sources and Limits
This guide is based on Peter's video and supplied transcript, Anthropic's Fable 5.2 model guidance, the Opus 5.5 model page, Peter's care-skill repository, and the NCI and ClinicalTrials.gov guides linked above. Checked 12 November 2026. Peter's video is sponsored by Anthropic. The prompt text above is from his video description; interfaces, connector access, limits and pricing can change. No email, calendar or medical data was accessed to write this article.