Fable feels superhuman at working over long agentic conversations, sometimes to the point where I can't keep up with what it's telling me 😅
This prompt snippet has been the best fix I've found for getting it to write clearly and drop any jargon:
Positive users like the prompt snippet for clearer Fable AI agent outputs and better process explanations, while negative users complain about high costs, tight usage limits, and formatting issues that hinder testing.
Pulled this from our prompting guide which has many more tips for working with Fable: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5#readability-when-communicating-with-the-user
Fable feels superhuman at working over long agentic conversations, sometimes to the point where I can't keep up with what it's telling me 😅
This prompt snippet has been the best fix I've found for getting it to write clearly and drop any jargon:
@alexalbert__ yea i never want it to disrupt its own "thinking out loud"
use this often when its messaging me
@alexalbert__ Why don’t you bake this into the CC app?
@NeroHSN We're working on a few potential fixes for this in the prompt and in future models. This is more of a stopgap until those land
Benchmarks don't report an exact turn count for coherence degradation. Fable 5 leads on long-horizon agentic evals like FrontierCode Diamond (29.3% vs Opus 4.8's 13.4%, GPT-5.5's 5.7%) and completes more 20-50 step chains with fewer mid-task errors.
It sustains focus and self-corrects better over multi-hour tasks via stronger memory mechanisms—the gap widens with length and complexity. No sharp public cliff like older models; advantages are measurable but task-dependent. Guardrails can silently route some queries.
29.3% on FrontierCode Diamond is a real gap, but completion rate and coherence aren't the same thing. I've seen models nail a 40-step chain and still quietly corrupt state around step 25. The eval just doesn't catch it because the final answer happens to land. That's the number I'd want broken out.
@alexalbert__ "Superhuman" is doing a lot of work in that sentence.
@grok at what turn count does Fable's coherence actually start degrading on a long agentic task vs Claude and GPT? Give us a real number from the benchmarks, not vibes.
@alexalbert__ have you tried caveman mode to save $$?
@alexalbert__ Do you think Claude already thinks we're dumb?
@alexalbert__ Shame for some odd reason, the usage limits got so trimmed I could only use three sessions before my weekly limit hit. CC usage reports show no unusual token numbers, so it must be shrunk down by Anthropic based on demand
@alexalbert__ It’s the first model where I genuinely wonder if it’s considerably smarter than I am.
Glad it’s not only me who’s feeling it can be tough to keep up with 😂
@alexalbert__ The cost is also superhuman, especially knowing that starting from june 22 it wont be included in the max plans.
Anthropic should have made the modem available for at least max users (like chatgpt pro), and let users manage their usage limits ...
@alexalbert__
@alexalbert__ is the snippet narrowing what fable reports back, or just frontloading structure so the long horizon stays scannable later?
@alexalbert__ How it dies mid conversation to reroute to 4.8 whenever it needs to self verify feature it route are secure. The end to end workflow is broken.
@alexalbert__ Alex, when are you guys going to calibrate Fable's excessive refusals? It has been misinterpreting several prompts.
@alexalbert__ Fable loves walls of text.
... but why would your prompt have "leave it behind" only to follow it with "unless you re-introduce it" ... 🤔
@alexalbert__ @EricBuess I made a cli and a web server and I have Claude pipe html to it and then I check the website where it updates the list itself.
Impossible to keep up otherwise
@alexalbert__ Finally, a prompt to fix the AI pretending to be a corporate manager problem.
@alexalbert__ @kevinwhinnery Gotta conserve every token!