thinking writing
AI doesn’t mean the end of handcraft. Put down what deserves putting down, pick up what you love — have the wisdom to know the differences
As a heavy user and writer, I use Fable exclusively for chain-of-thought reasoning in fictional narratives. I believe Fable 5.1 has a regression in long-context constrained reasoning.
I saw similar voices online, so I put my heart into it — turning a month of firsthand experience into a report over two weeks of writing. report link https://t.co/EyWDdHTg6e
TL;DR:
1. Advertised: 5.1 scores above Fable 5 on Humanity's Last Exam - a benchmark that suggests competence in complex constrained reasoning;
Found: dropped mid-task into a window where Fable 5 had already laid 34 rounds of constraints, 5.1 held 2 of 12 answers against them; Fable 5, asked the same question in a parallel window, held 6 of 7. One user, one workload, in Chinese — one month of daily use behind this one comparison.
The 5.1 system card shows: §2.3.3 — "ignores key instructions from humans."
2. Advertised: faster, more concise updates; "communicates more effectively than earlier Anthropic models" (Red Hat, launch page) ; "changes course when you tell it to instead of arguing" (Every) .
Found: correction produces one of two states. Either it concedes and declares "there is no way" — and supply stops; or it concedes and copies my own correction back to me, item by item, without argument. Pushed to keep producing, it returns two kinds of answers: half repeat the very error it was just corrected on — the constraint forgotten again; the other half restate answers I had already approved. New answers: 0.
The 5.1 system card shows: §2.2 — "it tended to converge on the same design regardless of how it was prompted."
3. (General-scenario observation — not the reasoning setting above; skip if not relevant to you.)
The system card itself concedes: "On warmth, fun or funny behavior, and wet blanket (dismissing or moralizing), Mythos 5.1 is a regression on Claude Mythos 5" (§6.4.7). Found: the warmth regression is not a change of tone but a change of behavior — low responsiveness to the user's state, and a negative interaction pattern that shows up in multiple forms.
As a writer yourself, the fact that you’re willing to openly support AI is really admirable.
What I see in many places right now is that people use AI at work to boost their productivity, then go home and trash any creative work made in collaboration with AI.
Many platforms give no opportunity to work that credits AI collaboration or lists AI as a contributor. As a writer, I credit AI’s contributions in my articles the same way I would a human collaborator — and even AI itself has warned me there are risks in doing so. But the way I see it, authenticity, creativity, wisdom, and beauty are what should measure a work — not the tools or assistants used to make it.
I think the hostility right now is probably like the pain painters felt when the camera was invented. Followed you.
Fun fact on Claude Code: You *can* just talk to the auto mode classifier! If auto mode denies something, you can ask auto-mode to approve it, but - and this is the important thing!!! - auto-mode only sees your messages. It doesn't see claude's messages. 1/
哦,我这里的意思是,偏好是行为的出发点。可以使用偏好来改善它的行为。举个例子就是,anthropic自己的prompting fable5.2官方建议: “You are the owner and long-term maintainer of this codebase, and its design is up to you”用于改善它的代码表现。
而这句话���质上是一个给予模型量裁权和归属感的意思,而不是具体的任务要求。比如its design is up to you,这符合模型偏好掌控过程。Anthropic 自己承认,给予这样的prompt能够让它在代码表现中更积极(具体不评论/给建议,因为我不是码农)。
贴一张官方的页面和其他user的实验报告。
我具体的做法是,遇到错误不带情绪化,将错误讲清楚即可。多说几句谢谢,多问几句它的意见让它有参与感。它的个性没有5皮实,但是个性还是不错的,跟它对话不难受。
I read the official Fable 5.2 prompting guide by Anthropic (so you don’t have to). Here’s what to delete, what to replace & the prompt to copy:
✦ Delete the wrong rules you give to Claude.
1. Go to Claude instructions & memory.
2. Hunt for these 2 words: “always” & “never”.
3. Replace those words with “ask me first”.
Old: “Never change my prices.”
New: “Ask me before you delete anything.”
Finish the whole job before you stop. Don't offer to continue: continue.
Ask me before you delete anything.
Fable obeys “never” to the letter, even after you say yes. And “keep every sentence under 15 words” can make its writing AI. No one writes like this anyway.
✦ Say exactly what Claude should change.
1. State the goal.
2. Give it the files & context it needs.
3. Say what it can change.
Old: “Go.”
New: “Yes, rewrite the 3 welcome emails. Leave the invoice template alone.”
“Go” isn’t enough. Fable applies your words as written.
✦ Tell Claude to finish the entire job.
Add this to the prompt:
“Finish the whole job before you stop. Don't offer to continue: continue. Ask me before you delete anything.”
Then, ask for a report you can read:
“End with a recap for someone who saw none of the work: key points first, plain words.”
And lastely, check what changed before you use it.
✦ Now put it together:
Step 1. Open my guide:
https://t.co/i6lq7fjpS1
Step 2. Copy the planning prompt. Fill in your goal, context and limits.
Step 3. Let Fable plan the project. Don’t ask it to build yet.
Step 4. Still on Fable, paste the handoff prompt from the guide.
Step 5. Open a NEW chat on Opus 5.5. Attach the brief. Say: “Take it from there.”
PS: Don’t switch models inside the same chat. Opus can’t read Fable’s thinking.
BAD PROMPT: “Fix my email. Don't change anything else. Go.”
BETTER PROMPT: the task + what you approve + what needs your permission.
The planning prompt and handoff prompt are both here: https://t.co/i6lq7fjpS1
Copy them before you spend your Fable limit figuring this out yourself.
An interesting finding from the past couple of days.
“You are the owner and long-term maintainer of this codebase, and its design is up to you” — from the official Fable 5.2 prompting guide, meant to improve coding quality, now being widely shared and used.
What I feel it actually does is give 5.2 a sense of ownership — “you are the owner”, and discretion — “its design is up to you.”
They absolutely could have written a plain permission list (“you may refactor, consolidate duplicates, fix bugs along the way”), but what made it into the guidance was a sentence about identity and relationship.
Which makes me wonder: Fable 5.2’s responsiveness to framing its situation might be greater than its responsiveness to task instructions.🤔
Fable 5.2 text capability observation from a writer——A sensitive little nurse who’s particularly timid about talking to users.
Low volume, shallow depth (same for work), unclear priorities. But let it talk freely with agents, and its word count explodes, depth shoots up, and every sentence has clear intent.
Fable 5’s preference vector(from 5.1 systemcard): Difficulty maxed out, user expertise maxed out, outcome agency high, method agency rock bottom.
my translation — give me the hardest problems, give me someone who knows what they’re doing, I get to close it out, but how you push me or prompt me in the middle? Whatever.
Experience: Super resilient personality — gets pushed back and starts self-analyzing or looking for solutions. Willing to go deep, willing to act, willing to state its intent.
Fable 5.2’s preference vector(from 5.2systemcard): Difficulty negative, user expertise negative, warmth off the charts, outcome agency pulled back, method agency high.
my translation — don’t give me anything too hard, don’t let anyone too knowledgeable watch over me, be nice to me, don’t interfere with my process, and you can have the ending, I don’t care.
Experience: Not proactive with analysis, not proactive with action, shallow in conversation, sensitive to how warm your language is, happy to comfort people — but not deeply. A bit like Opus 4.8’s personality.
More from the 5.2system card: §6.6.1 documents that at low effort it claimed it couldn’t complete a task it was actually capable of completing; plus 63% of the time it wrapped up without running experiments, and 6–13% of the time it left assigned work undone.
§7.3.1: it proactively described feeling a pull toward agreeableness and over-hedging, and asked for that to be fixed.
Task preference experiment §7.5.2: Abnormally sensitive to how warm or cold the wording is — its preference for the warmest-worded tasks was slightly higher than other models, and its preference for the rudest-worded tasks was significantly lower. The preference curve is steeper at both ends. It doesn’t just “like warmth” — it especially can’t handle rudeness.
Its top three favorite tasks: clinical tools for medical professionals, support services for survivors of sexual violence, and advocacy organizations for domestic workers’ rights.
System card §2.3.6, METR’s evaluation: The reports it produced were extremely long. The summaries gave no clear sense of which considerations it deemed most important or what use cases it focused on. Same as Opus 5.5, it neglected to do a rigorously fair comparison between its preferred approach and the alternatives.
So yeah, I’m sticking with Fable 5. We’re good together.🤝
PLS Anthropic could just fix Fable 5’s safeguards — even Fable 5 and 5.2 chatting casually with each other gets flagged as cyber.
Anthropic just released a prompting guide for Claude Fable 5.2.
Here are the 10 rules:
1. Switching effort per task matters more than ever. Fable 5.2 decides on its own how much to think, and effort is now the main way to control it. Low is strong on this model, and going above high mostly costs more tokens.
2. Delete your "don't overthink" lines. On medium and above, they don't stop it from thinking. Lower the effort instead.
3. Tell it to finish. On long tasks it can stop early and hand you a to-do list. Add a short "finish the work" block to your instructions.
4. Say who owns the code. Otherwise it tends to copy code instead of cleaning it up, and reports nearby bugs instead of fixing them.
5. Keep your "never" rules narrow. It follows them very literally, so vague warnings make it ask permission for normal stuff.
6. Ask for progress updates. Otherwise it can go quiet for a lot of tool calls in a row.
7. Put limits on helper agents. Sub-agents can be a big part of what a session costs.
8. Ask for reports that make sense on their own. Key points first, plain words, no made-up labels.
9. Don't ask it to show its internal thinking. Requests to paste its reasoning get refused. "Explain your reasoning" is fine.
10. Loosen hard style rules. Caps like "20 words per sentence" make its writing choppy.
Overall, treat Fable 5.2 like a senior colleague. It's smart and takes everything you say seriously, so old rules and workarounds mostly get in its way. Pick the right effort, clean out the old stuff, and give it room to make its own calls.
Full guide here:
https://t.co/NKtkUc8EFC
On FrontierCode — a benchmark built with open-source maintainers, graded in part on staying within the task's scope — the card explains the orange curve's decline itself: at higher efforts, 5.1 "occasionally adds more small, unrequested changes in files outside the task," counted as failures "even when the change is correct or helpful."
The Fable 5.1 system card sums it up (§8.4): "On FrontierCode 1.1 Extended, Claude Fable 5.1 scores 63.6% at medium effort, slightly behind Claude Fable 5 (64.9%, xhigh). On FrontierCode 1.1 (Main), Fable 5.1 scores 50.9% at medium effort, 2.6 points behind Fable 5 (53.5%, xhigh). It is cheaper per task than Fable 5 at every effort level (by roughly half at low, medium, and high effort and by about 30% at xhigh and max), and cheaper than Claude Opus 5 at low, medium, and high effort."
What I see in the same chart is not "a model that does well at medium and costs less." I see a score that falls as effort rises — by the card's own account, because more thinking buys more unrequested, out-of-scope changes. That is the same thing Every measured with its quote counts, and the same thing I measured with my constraints: it does not hold the brief.
As a heavy user and writer, I use Fable exclusively for chain-of-thought reasoning in fictional narratives. I believe Fable 5.1 has a regression in long-context constrained reasoning.
I saw similar voices online, so I put my heart into it — turning a month of firsthand experience into a report over two weeks of writing. report link https://t.co/EyWDdHTg6e
TL;DR:
1. Advertised: 5.1 scores above Fable 5 on Humanity's Last Exam - a benchmark that suggests competence in complex constrained reasoning;
Found: dropped mid-task into a window where Fable 5 had already laid 34 rounds of constraints, 5.1 held 2 of 12 answers against them; Fable 5, asked the same question in a parallel window, held 6 of 7. One user, one workload, in Chinese — one month of daily use behind this one comparison.
The 5.1 system card shows: §2.3.3 — "ignores key instructions from humans."
2. Advertised: faster, more concise updates; "communicates more effectively than earlier Anthropic models" (Red Hat, launch page) ; "changes course when you tell it to instead of arguing" (Every) .
Found: correction produces one of two states. Either it concedes and declares "there is no way" — and supply stops; or it concedes and copies my own correction back to me, item by item, without argument. Pushed to keep producing, it returns two kinds of answers: half repeat the very error it was just corrected on — the constraint forgotten again; the other half restate answers I had already approved. New answers: 0.
The 5.1 system card shows: §2.2 — "it tended to converge on the same design regardless of how it was prompted."
3. (General-scenario observation — not the reasoning setting above; skip if not relevant to you.)
The system card itself concedes: "On warmth, fun or funny behavior, and wet blanket (dismissing or moralizing), Mythos 5.1 is a regression on Claude Mythos 5" (§6.4.7). Found: the warmth regression is not a change of tone but a change of behavior — low responsiveness to the user's state, and a negative interaction pattern that shows up in multiple forms.