Fable 5.2 handling manual: Anthropic's prompting guide, TL;DR
Anthropic published a full guide on how to prompt Claude Fable 5.2. It's long, and it's written for developers building on the API.
Most of it comes down to one idea: Fable 5.2 does exactly what you write. Habits that worked on earlier models now backfire, and a few new ones pay off.
Here are the 15 rules that matter if you use Fable in Claude or Claude Code.
🧭 Picking the model
1. Opus 5.5 for daily work, Fable for the big job.
Fable costs twice as much on the API ($8/$40 vs $4/$20 per million tokens). Anthropic itself says Opus 5.5 stays the default. Reach for Fable when Opus at higher effort still falls short.
2. Don't switch models mid-chat.
Caches are per model, so the new model rereads the whole chat at full price. And Opus can't read Fable's thinking, so that reasoning is gone. Start a fresh chat instead.
3. Let Fable plan and cheaper models build.
Early testers had good results with Fable as the planner and orchestrator for other models. Have it write the plan and a handoff brief (goal, decisions, open questions, a checklist), then run the brief in a new Opus chat.
✍️ Writing prompts
4. Swap "never" for "ask me first, because...".
Under a "never" rule, Fable took a step the person in charge had approved about 1 time in 10. When the rule said to get a go-ahead first, 9 times in 10.
Old: "Never run migrations."
New: "Ask me before running migrations: the client's data lives in that database."
5. Name what's approved.
A bare "go" leaves it guessing. Fable applies approval rules as written, so spell out the scope: "Yes, refactor the checkout views. Leave the payment code alone."
6. Tell it to finish.
On long tasks Fable sometimes stops at "Want me to continue?", more often at Low and Medium effort. A "finish the whole job, don't offer to continue" block halved the sessions that ended with doable work left, at about 30% more output tokens. It also asks less with it, so keep your own confirmation step for anything risky.
7. Decide who owns the code.
Say "this codebase is yours" and Fable restructures as it goes. It merged repeated code in about half the tries (under 1 in 10 without), used 1.5x the tokens and made changes nobody asked for. Leave the line out when you want minimal changes. When you use it, review the diff, not the summary: the summary often skips small changes.
8. Ask for reports a stranger can read.
After long runs, Fable's reports get dense. Anthropic suggests asking for reports "mostly in the style of ASD-STE100 Simplified Technical English", the plain English made for aircraft maintenance manuals. Key points first, short sentences, no made-up terms.
9. Set length in the prompt, not with effort.
Effort barely changes how long replies are.
🎚️ Effort
10. Start at High.
Most of the capability gain is already there. Above it, gains are uneven and tokens pile up. Use XHigh only where you've measured a gain. Max is for ceiling tests.
11. Lower effort instead of writing "don't overthink".
Those lines stopped working: at Medium and up, Fable thinks anyway. At High it thinks before almost every reply, even "hi". Drop to Medium for planning and easy turns.
🧹 Cleaning up old instructions
12. Clear out stale notes and memory.
Cautions pile up, and Fable follows them closely. With a dozen notes from past sessions in context, one of them saying never to retry, retries after a busy tool fell from nearly every try to about 4 in 10.
13. Name the steps that need no approval.
A line listing retries among the steps that need no go-ahead raised them from almost none to about half the tries. A general "act more freely" changed nothing.
14. Delete hard style rules and token budgets.
Fable follows "one instruction per sentence, 20 words or fewer" so closely the prose turns choppy. A token budget written into the prompt changes how much Fable spends: removing it cut cost on coding tasks at the same quality.
🤖 Claude Code
15. Fable conducts, Opus workers execute.
Run Fable 5.2 as the main model and your subagents on Opus 5.5. Fable plans and checks the work, the workers do the tasks. Watch the bill: subagents can be a large share of a session's cost, and at XHigh and Max Fable starts extra ones to review its own work.
My own agent's rules file has 18 "never"s. Rule #4 hit close to home 😅
Which of these surprised you? And what's the oldest line still sitting in your CLAUDE.md?
If this saves you some tokens, share it with someone who runs Fable all day. Thanks and take care!
3
1
166
Fable 5.2 effort, per Anthropic: start at High 🎚️ Most of the capability gain is already there. Above it, gains are uneven and tokens pile up.
XHigh only where you've measured a gain. Max for ceiling tests 🧪
What do you run by default, and why?
1
44
📜 Fable 5.2 follows old prompt rules to the letter.
Got "one instruction per sentence, 20 words or fewer" in there? It obeys so closely its prose turns choppy 🪓
The fix: delete the style rules you no longer need.
37
Buried in Anthropic's Fable 5.2 guide: ask for reports "mostly in the style of ASD-STE100 Simplified Technical English".
That's the plain English made for aircraft maintenance manuals ✈️ Short sentences, one idea each, active voice.
Would you read agent reports like that? 🤓
30
The split early testers liked, per Anthropic: Fable 5.2 plans 🧠, cheaper models build 🔨
Fable costs 2x Opus 5.5 on the API ($8/$40 vs $4/$20 per M tokens). Spend it on the thinking, then hand Opus a brief and a checklist.
Which model does your planning? What's your split?
27
Fable 5.2 applies your words as written. So a bare "go" leaves it guessing what you meant 🤷
Anthropic's advice: name what's approved ✅
"Yes, refactor the checkout views. Leave the payment code alone."
What's the weirdest "go ahead" you've ever given an agent?
35
Fable 5.2 sometimes stops at "Want me to continue?" right after you walk away 🚶
Anthropic's fix, a "finish the whole job" block in the prompt, halved the sessions that ended with doable work left. The cost: ~30% more output tokens 🧾
26
"Answer directly, don't overthink" stopped working on Fable 5.2. At Medium effort and up, it thinks anyway 🤔
At High it thinks before almost every reply. Even "hi" 👋
The dial is effort, not the prompt. Which lines in your prompts are dead weight now?
38
Switching from Fable to Opus mid-chat costs more than it looks 💸
Caches are per model, so Opus rereads the whole chat at full price. And it can't read Fable's thinking, so that reasoning is gone.
Start a new chat with a handoff brief 📝 Do you switch mid-chat?
23
Tell Fable 5.2 "this codebase is yours" 🔑 and it starts acting like it.
In Anthropic's tests it merged repeated code in about half the tries (under 1 in 10 without), used 1.5x the tokens, and made changes nobody asked for 😬
Would you give an agent that much rope?
33
Your agent's old notes are making it timid 🫣
Anthropic put a dozen notes from past sessions in context, one saying never to retry. Retries after a busy tool fell from nearly every try to about 4 in 10.
When did you last clean out your agent's memory? 🧹
25
Fable 5.2 obeys "never" even after you say yes.
❌ Rule: "never change prices." Then you say "ok, change this one." In Anthropic's tests of this pattern, it changed the price after your ok only 1 time in 10.
✅ Rule: "ask me before changing prices." Same ok, and now Fable does what you asked 9 times in 10.
My agent's rules file has 18 "never"s. Oops 😅
How many are in your CLAUDE.md?
1
65
The founder of the well-known WordPress plugin Yoast (I'm sure you've heard of it even if you've never used WordPress) has released a very handy Website Spec—a checklist of features that every good website should include. What's interesting about the checklist is that it's written for both humans and agents 👇


