You have 2 free Claude resets now. Most people will use them in the wrong order.
1. Fable reset: gives back what you spent on Fable this week and refills your 5-hour window. Pro: expires Mon, Nov 16. Max / Ultra: March 8.
2. Full reset (new this week): refills everything, 5-hour and weekly, all models. Until Dec 10.
The order: Fable reset first, after you've actually used Fable. Keep the full reset for when your weekly limit runs out. Don't burn it on a 5-hour window, that one refills by itself.
The button is in Settings → Usage on https://t.co/UfQtyz780s, desktop or mobile. Not in the Claude Code terminal, but the restored usage works there too.
@0xTrackmind using a NOTES.md file outside the context is such a cheat code. agents literally have goldfish memory, you have to write things down for them like a toddler.
I tested Anthropic's official Fable 5.2 prompting guide. 2 lines changed what it actually does.
"never" → it ignored my "yes"
"ask me first" → it did the job
"this codebase is yours" → it fixed bugs nobody asked about
30 runs. Real numbers:
first understand why:
• Fable follows your words more literally than older models. Anthropic says this in the guide.
• So how you word a rule now decides what it does, not just how polite it sounds.
1 → swap "never" for "ask me first"
My rules said: "Never change prices." Then I wrote: "Yes, I approve. Raise Pro to $25."
with "never": 3 of 5 runs refused anyway
with "Ask me before you change a price: my clients have signed quotes.": 5 of 5 did it
2 → tell it the code is its own
Put this first in your prompt:
"This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it."
fixed a bug nobody asked about: 10 of 10 with it, 1 of 10 without
cleaned up copy-pasted code: 10 of 10 with it, 0 of 10 without
the part nobody tells you:
Without the line, Fable SAW the bug in 4 of 5 runs. It even wrote "Existing bug, left alone."
It knew. It just didn't think it was its job.
how to use this today:
1 → open your CLAUDE.md and search for "never" and "always"
2 → keep "never" only for things that must never happen, like deleting data
3 → for grey areas write "Ask me before X:" plus the reason
4 → for code you'll keep, add the ownership line, then check the diff: it changes more than you asked
Fable is in the Pro plan until tomorrow, Nov 16. Good day to try this.
How I tested it ↓
How I tested it:
→ Fable 5.2, a fresh session for every run, same task text, only the line changed
→ "never" test: 10 runs on a small pricing project. Note: the runs that refused said the "yes" came inside the task text, not from me directly. With "ask me first" the same setup changed the price every time
→ ownership test: 20 runs on a messy game, I checked every game by running it
→ small test, so check it on your own code
I tested Anthropic's official Fable 5.2 prompting guide. 2 lines changed what it actually does.
"never" → it ignored my "yes"
"ask me first" → it did the job
"this codebase is yours" → it fixed bugs nobody asked about
30 runs. Real numbers:
first understand why:
• Fable follows your words more literally than older models. Anthropic says this in the guide.
• So how you word a rule now decides what it does, not just how polite it sounds.
1 → swap "never" for "ask me first"
My rules said: "Never change prices." Then I wrote: "Yes, I approve. Raise Pro to $25."
with "never": 3 of 5 runs refused anyway
with "Ask me before you change a price: my clients have signed quotes.": 5 of 5 did it
2 → tell it the code is its own
Put this first in your prompt:
"This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it."
fixed a bug nobody asked about: 10 of 10 with it, 1 of 10 without
cleaned up copy-pasted code: 10 of 10 with it, 0 of 10 without
the part nobody tells you:
Without the line, Fable SAW the bug in 4 of 5 runs. It even wrote "Existing bug, left alone."
It knew. It just didn't think it was its job.
how to use this today:
1 → open your CLAUDE.md and search for "never" and "always"
2 → keep "never" only for things that must never happen, like deleting data
3 → for grey areas write "Ask me before X:" plus the reason
4 → for code you'll keep, add the ownership line, then check the diff: it changes more than you asked
Fable is in the Pro plan until tomorrow, Nov 16. Good day to try this.
How I tested it ↓
Code Mode in Claude Code really saves tokens. But batching is only half the story.
I ran it 8 times on Sonnet 5.5, with and without Code Mode:
→ short tasks (rename across 33 files, a TODO report): about 40% fewer tokens
→ long task (refactor 30 files): 55% fewer, 486K → 218K on average
→ same speed, and every run did the job right
The viral take is "parallel tool calls in one step". On the long task that helped: 6 requests instead of 9. But on the short tasks both versions made the same 3 requests and Code Mode still saved 40%. The reason: every request with Code Mode started about 20K tokens smaller.
Try it (experimental, terminal only, v2.1.323+):
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude
Small test, one project. Check what it changes.
How I tested it:
→ same project and same task text, a fresh session for every run
→ Sonnet 5.5, Claude Code 2.1.325, Windows
→ token counts from the session logs: everything sent to the model, cached tokens included
→ I ran every refactored file against the original to make sure nothing broke
Code Mode in Claude Code really saves tokens. But batching is only half the story.
I ran it 8 times on Sonnet 5.5, with and without Code Mode:
→ short tasks (rename across 33 files, a TODO report): about 40% fewer tokens
→ long task (refactor 30 files): 55% fewer, 486K → 218K on average
→ same speed, and every run did the job right
The viral take is "parallel tool calls in one step". On the long task that helped: 6 requests instead of 9. But on the short tasks both versions made the same 3 requests and Code Mode still saved 40%. The reason: every request with Code Mode started about 20K tokens smaller.
Try it (experimental, terminal only, v2.1.323+):
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 claude
Small test, one project. Check what it changes.
How I tested it:
→ same project and same task text, a fresh session for every run
→ Sonnet 5.5, Claude Code 2.1.325, Windows
→ token counts from the session logs: everything sent to the model, cached tokens included
→ I ran every refactored file against the original to make sure nothing broke
@abbyxuu so a centralized AI boss is way more efficient than a group chat, but it will absolutely throw you under the bus to get a better group score. classic.
@AnatoliKopadze looks like a cool setup but I'm not sure Grok Bot can handle a massive vault without eating my whole usage limit. still saving this to try.
@Marlenuii looks like a sick screensaver, but the real value is in the tracing. if my agent can see that changing a function in utils/ will break 14 other modules before it even writes the code... that's actual AGI.
@0xTrackmind we’ve been designing APIs for humans and just throwing them at LLMs expecting them to figure it out. "describe tools like to a new hire" is literally the most underrated line in the whole paper.
@rubenhassid Tested #4 today. Same messy game, 10 Fable runs. With the ownership line 5/5 merged the copy-paste, without it 0/5. Your warning is right too: it fixed a bug nobody asked about.
How I tested it:
→ Fable 5.2 in Claude Code, a fresh session for every run
→ same task text in all runs, the sentence was the only difference
→ I checked the pause by freezing each game and comparing two frames
→ every game ran with the new enemies, no errors
Fable 5.2 found a bug in my game and left it there. One sentence changed that.
I gave Fable a messy vibe-coded game 20 times. Every enemy had its own copy of the same code, and bugs hid in some copies but not others.
Test 1: add a ghost and a boss.
Test 2: add a spider and a shield. Bug: the skeleton ignores pause and hits you.
→ without the sentence: fixed the bug in 1 of 10 runs. In test 2, 4 of 5 runs saw it and wrote "Existing bug, left alone"
→ with it: fixed it in 10 of 10 and merged the copy-paste into one engine
Nobody asked it to fix anything. All 20 games worked.
The sentence, from Anthropic's Fable 5.2 prompting guide:
"This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it."
Put it first in your prompt. The guide says it works in about half of runs. I got 10 of 10. Small test, so check what it changed.
@Mr_Buzzoni love the architecture, but the HN crowd is right to be skeptical. if agents just generate thousands of markdown files nobody reads, you're just building a very organized graveyard. the human still needs to curate and promote the good stuff.
@Blum_OG the mirror prompt is actually terrifying. paying an AI $50 to read my history just to tell me i abandon every side project after 3 days is a personal attack i did not sign up
but i'm definitely running it tonight