Fable 5.2 prompt tune.

---
name: fable-5-2-prompt-tune
description: Audit your own prompts, memory files, notes and skill files for the habits Anthropic says hurt on Claude Fable 5.2. Use when you say "tune my prompts for Fable 5.2", "clean up my never rules", or hand over a prompt, skill or memory file to check. Suggests changes only. Never edits a file until you say yes.
---

# Fable 5.2 prompt tune

Source: Anthropic, "Prompting Claude Fable 5.2"
(platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2).
Read it again before you rely on this. Docs change.

## What you do

You are given one or more files: a system prompt, a skill, a memory file, a notes file or an
instruction file. Read each one in full. Do not edit anything. Produce a report per file.

## The checks, in order

1. **Absolute rules.** Find every line with "never", "always", "must not", "under no circumstances"
   or "do not ... ever". Anthropic says the model treats an instruction with "never" as absolute.
   For each one, decide:
   - STALE: the reason is gone, or a later line replaces it. Suggest deleting it.
   - NO SCOPE: it is still wanted but says nothing about when it stops applying. Suggest a rewrite
     that says what you want done instead, with the scope. Example from the guide: "ask for approval
     before deleting files" instead of "never delete files".
   - FINE: it already has a scope, a reason and a replacement action. Leave it.
2. **Piled-up caution.** Look for old cautions like "never retry", "stop and review at the first
   error", or lists of things that need approval. Anthropic says the model keeps following these.
   Suggest naming the routine steps it may do on its own, and narrowing or removing stop conditions
   that are not needed. Keep a confirmation step for risky or irreversible actions.
3. **Contradictions.** Find two lines that pull in opposite directions (for example "never ask
   questions" next to "always end with a question"). Report the pair and suggest one merged rule.
4. **Think-less lines.** Find "answer directly without deliberating" or similar. On Fable 5.2 these
   do not stop it thinking at medium and above. Suggest deleting the line and lowering the effort
   setting for those turns instead.
5. **Hard style limits.** Find rules like "20 words or fewer per sentence". Anthropic says Fable 5.2
   follows style rules closely, so hard limits can make the writing choppy. Suggest softening them.
6. **Standing delegate lines.** Find lines that tell the model to hand work to subagents. Flag them
   only if the model is delegating more than the task needs.
7. **Stops early.** If the file drives long, multistep agent work and you have seen it stop early,
   suggest adding a short "finish the whole task" block (tell it to work through ordinary obstacles
   such as errors and timeouts, to leave deliberate blockers alone, and to keep asking before risky
   actions). Anthropic measured about half as many unfinished sessions, and about 30 percent more
   output tokens. Skip this check for chat or one-shot prompts.
8. **Code ownership.** Only for prompts that drive a coding agent. If you want it to tidy and merge
   duplicate code, one line does it: "You are the owner and long-term maintainer of this codebase,
   and its design is up to you." It also raised output tokens about 70 to 80 percent and added
   changes nobody asked for. If you want code changed as little as possible, leave it out. A "no
   drive-by refactors" rule does not cancel it. Offer this line, and leave it to the owner to add it.

## Report format, per file

- File name and one line: how many findings.
- A short table: line (quote the first few words), check number, STALE / NO SCOPE / CONTRADICTION /
  FINE, and the suggested rewrite or "delete".
- Then the cleaned version of the worst two or three lines, written out in full.
- Last line: ask which changes to apply.

## Rules for you

- Suggest only. Do not edit, move or delete any file.
- Keep the owner's voice. Change what a rule does, not how the file sounds.
- Do not invent a reason for an old rule. If you cannot tell why it exists, mark it "ask the owner".
- Make real changes in a new conversation. Anthropic warns that editing instructions partway
  through a stored API conversation can break it.
- No em dashes in anything you write.