Migrate to Fable 5.2
Migrate to Fable 5.2 (migrate-to-fable-5-2) is a Claude Code skill that moves prompts and agent setups to Claude Fable 5.2: it follows Anthropic's prompting guide to recalibrate effort, remove instructions written for earlier models, and add one fix for each behavior you observed after switching.
npx skills add lior-pesoa/skills --skill migrate-to-fable-5-2View on GitHub ↗Raw for agents: /library/migrate-to-fable-5-2.md
Anthropic published a guide to prompting Claude Fable 5.2 (opens in a new tab). It lists what to change when you move a prompt, a skill or a whole agent setup over from Claude Fable 5.1 or Claude Opus 5.5. I turned it into a skill, so the agent does the checking.
The API changes in a new model return a 400, so you find them. The changes in this guide don't raise errors. An effort level tuned on another model, a system prompt line that stopped working, a memory file full of old cautions: they show up as cost, slow turns and worse results.
What changes with Fable 5.2
These are the points from Anthropic's guide that the skill checks. The numbers come from Anthropic's own testing, as the guide reports them.
- Effort doesn't carry over. The levels rank the same way across models, but the amount of thinking at each level differs. The guide says to set
highexplicitly, then testmediumandxhighon your own evals. - "Answer directly without deliberating" no longer stops the model from thinking at
mediumand above. Lower the effort for those turns. - The model follows instructions closely and treats "never" as absolute. Cautions that piled up in memory and notes make it ask for approvals nobody needs.
- A token budget shown in the prompt changes how much the model spends. On coding tasks, removing a large one lowered cost with the same results.
- It pays attention to who owns the code. One sentence saying the codebase is its own took consolidating duplicated code from under 10% of runs to about 75%, and raised output tokens by 70 to 80%.
- On long tasks it sometimes ends a turn before the work is done, more often at
lowandmedium. With the guide's finishing text, about half as many sessions ended that way, for about 30% more output tokens. - Updates between tool calls come back as thinking blocks that are empty by default, so a client that shows only text looks silent.
- A prompt that asks the model to write out its reasoning can be refused, under the
reasoning_extractioncategory.
What it does
- Finds every place instructions reach the model: system prompts,
CLAUDE.md, skills, agent files, memory, notes, tool descriptions and the code that builds requests. - Sets
efforttohighexplicitly and plans a sweep atmediumandxhigh. It says so when no eval has confirmed a level. - Checks for nine kinds of leftover from earlier models, each with search terms, the reason it misbehaves now and what to do about it.
- Matches each behavior you saw on Fable 5.2 to the fix from Anthropic's guide, with the suggested text, its measured effect and its token cost.
- Asks before it removes anything it can't judge, like a caution whose reason may still hold, and leaves your confirmation step for risky actions alone.
- Writes a report: findings with file and line, what changed, what it left alone and what to measure next.
When to use it
Use it when you move an agent, a prompt or a Claude Code setup to Claude Fable 5.2 or Claude Mythos 5.2, or when an agent got slower, more expensive or more hesitant after the switch. It covers prompts and behavior. The API changes that return errors are in Anthropic's migration guide.
How it works
- 01Reads the repo, then asks what only you know: the model you're moving from, whether someone watches the agent work, what you've seen on Fable 5.2, and whether to edit or only report.
- 02Lists every file and code path that carries instructions into the model's context.
- 03Calibrates effort before it touches any prompt text.
- 04Audits the list for instructions written for earlier models. Removing comes before adding.
- 05Adds one fix per behavior you observed. If you haven't run on Fable 5.2 yet, it gives you a list of what to watch for and adds nothing on a guess.
- 06Reports, and reminds you to ship prompt changes to new conversations only, because a stored conversation resumed with a new system prompt can return a 400.
Questions
Where does the guidance come from?
From Anthropic's page Prompting Claude Fable 5.2 (opens in a new tab), as it read on 12 November 2026. The prompt texts in the skill are quoted from it. When the agent can fetch web pages, the skill tells it to prefer the live page where the two differ.
Does it change my API code?
No. Model IDs, the computer use tool, advisor pairing and the other changes that return a 400 are in Anthropic's migration guide (opens in a new tab). In Claude Code, /claude-api migrate this project to claude-fable-5-2 handles them. This skill covers what doesn't raise an error.
Why doesn't it add all of Anthropic's suggested prompts?
Each one costs tokens, and the guide says to add them one at a time, for a behavior you observed. In Anthropic's testing the finishing text raised output tokens by about 30%, the progress-update text by about 20% and the code-ownership sentence by 70 to 80%. Added together up front, they make the agent slower and hide which change did what.
I only use Claude Code, not the API. Is it still useful?
Yes. Claude Code builds the requests for you, so the skill skips the harness checks and audits what you control: CLAUDE.md, skills, agent files, memory and notes.
Has it been tested?
Not on a real migration yet. Version 0.1.0 is written straight from Anthropic's guide.
How do I install it?
Run npx skills add lior-pesoa/skills --skill migrate-to-fable-5-2, or copy the skills/migrate-to-fable-5-2 folder into ~/.claude/skills/. Then type /migrate-to-fable-5-2, or tell the agent what changed since you switched.