TL;DR The Fable 5.1 prompting guide has sixteen sections. Six are already inside Claude Code's system prompt as of v2.1.257, and only when the model is Fable 5.1. Six aren't the Claude Code user's job, and the effort one is set with
/model, not with text. The three left over are yours: targeted edits, scope and tests, and search before answering at low effort. Paste those intoCLAUDE.mdand delete whatever argues with them.
The day Fable 5.1 shipped, Anthropic also published its prompting guide: a list of behaviors that changed from Fable 5 and, for each one, the paragraph to drop into your system prompt to correct it. It is written for people building against the API with their own harness. That is the trap if you read it from Claude Code: the harness is Anthropic's, and it read the guide before you did.
If you're after the price and the minimum version, that's in the launch tip. This is the next step: what to do with the guide once 5.1 is in front of you. Opus 5 had the same moment, and half of that advice was about what to remove.
What Claude Code already carries
The model catalog ships inside the binary, and the Fable 5.1 entry has a capability Fable 5 lacks: fable_5_1_prompt_bundle. With it on, Claude Code injects the guide's blocks on its own. You can watch it happen by asking the model itself, same binary, only the alias changed:
$ claude -p --model claude-fable-5-1 "Does your system prompt contain a
sentence beginning with 'Before you start, say in a line'? And a section
titled 'Writing for the user'?"
YES. "Before you start, say in a line what you're about to do; brief updates
while you work help the user follow along."
Section titled "Writing for the user": YES.
$ claude -p --model claude-fable-5 "<same question>"
NO. The closest sentence begins "Before your first tool call, say in a
sentence what you're about to do", which is different wording.
NO. The section covering that topic is titled "Communicating with the user".
That first sentence is the guide's, word for word. Fable 5, on the same binary, gets the older wording. Here is what changes hands when the model is 5.1:
| Guide section | In Claude Code with Fable 5.1 |
|---|---|
| Ask for user-facing progress updates | The guide's sentence, verbatim, in the system prompt |
| Batch independent tool calls | The guide's nudge, verbatim, appended after every tool result |
| Command output the user never sees | The guide's note, verbatim, after every Bash call with long output |
| Finish the whole task | Verbatim, though Fable 5 already had this one |
| Scope of the deliverable | Its own rewritten block, which covers scope but says nothing about tests |
| Writing density / formatting | Its own block, "Writing for the user" |
Plus two mechanisms the guide describes for harness builders that Claude Code already has: the lead keeps working while subagents run, and a reminder fires when the model has gone several turns without telling you anything.
This lands directly on your CLAUDE.md. If you have rules written against last year's models ("don't narrate, just give me the result", "no lists, no bold"), they now fight the injected blocks. The guide says so in as many words: remove them before adding anything. And if you keep a copy of the blocks above, you pay for them twice on every turn. In yesterday's tip I handed you the batching line for CLAUDE.md; as of 2.1.257 it's redundant, the binary sends it by itself.
grep -rns -i "don't narrate\|just the result\|no lists\|no bold\|hold all findings" \
~/.claude/CLAUDE.md ./CLAUDE.md ./.claude/agents/ ./.claude/skills/
The three lines that are missing
The three sections the binary doesn't carry are exactly the three the docs flag as raising output tokens on 5.1.
1. Targeted edits
5.1 tends to rewrite the whole file for a two-line change. The result is usually identical, the bill is not: you pay for the entire file in output. The guide says this instruction brings it back to Fable 5's behavior on small and medium changes.
The number of tokens used to edit files is best minimized, all else being
equal. When it will not affect the end result, surgically edit a file rather
than rewrite the entire thing.
2. Scope and tests
5.1 delivers what's asked and sometimes more: it fixes nearby code, extends behavior you never mentioned, and commits more test files than the task warrants. The scope block Claude Code already injects covers the first half but says nothing about tests. Anthropic measures that with this instruction the extras drop with no change in task success.
If, while working or testing, you find a pre-existing bug, a performance
concern, or behavior the task doesn't mention, don't fix, optimize or extend it
in this change; report it as a follow-up. Commit tests only where the task asks
for them or this repository already keeps tests for this kind of change, sized
like the neighboring test files. Don't turn scratch checks into permanent tests.
3. Search before answering at low effort
At low, 5.1 calls search less and answers from memory more. Ask it about anything that moves monthly (versions, models, tools) and you get stale data delivered with the same confidence. The guide's alternative is raising effort just for those turns; the line is for when you'd rather leave effort alone.
When a query centers on a name from a fast-moving area like AI models and
developer tools, the name itself is the thing to verify: search before
answering, even when you have some background on it. Familiarity is not a
reason to skip the search.
All three are short on purpose. A CLAUDE.md that grows with every launch ends up full of rules for models you no longer use, and this guide is the best case for pruning it: most of what the previous one asked for now ships in the binary.
The switch, if you'd rather it stayed quiet
Each new block comes with its own environment variable. They're undocumented, so treat them accordingly, but they work: with CLAUDE_CODE_TURN_UPDATES=false Fable 5.1 goes back to Fable 5's wording on progress updates. I ran the same question as above and the model quoted the old sentence. If what you want is less text between tool calls, try Concise mode first, which is the documented route.
What lands on neither list because it isn't the Claude Code user's job: append-only history (the binary handles it), the compaction instruction (it has its own), the quotation example, long outputs at xhigh and max, the image-cropping tools, and the phrasings that trigger classifier false positives.
Official docs: Prompting Claude Fable 5.1 · What's new in Claude Fable 5.1
Requirements
- Claude Code v2.1.257 or later with Fable 5.1 active (
/model fable). Checked on v2.1.258. - With Fable 5 or Opus 5 on the same binary, the new blocks are not injected: the gate is the model, not the version.