Claude Fable 5 is back, and right now it is the most capable model you can use.
You don’t have long with it though. It’s included in your Claude plan until 7 July, then it moves to pay-as-you-go credits.
While you have it, don’t run it like a smaller, older Claude with a bigger context window. It’s more powerful, but that’s not the point. The point is loops.
One of them ran on this very edition.
It built a rubric from my 110 past editions, caught this draft at 51/100, and rewrote it until it passed. You’re reading the pass.
In this issue:
The exact method to run your first loop today
The eval loop I use to make Claude check its own work
The rubric that scored this very edition, built on my own data
If you’re new here, welcome. If you’ve been reading for a while, thank you. Your subscription is what makes issues like this possible.
What is a “loop”?
Every loop has the same four beats.
Goal. What “done” looks like, written down before anything runs.
Act. The model takes one step towards the goal.
Check. Something compares the result against the goal.
Repeat. If the check fails, go again. If it passes, stop.
A prompt is one turn. A loop runs until the check passes. The check is the beat everyone skips, and I’ll keep coming back to it.
How to prompt Fable 5
Fable 5 isn’t a faster model. It’s a different one, and it needs different prompting.
Anthropic published the full guide. The most useful bits are paste-able. Here are the ones I actually use:
Give it the reason, not just the task.
Fable works better when it knows the purpose. Ten extra words, far more output.
I’m working on [LARGER TASK] for [WHO IT’S FOR]. They need [WHAT THE OUTPUT ENABLES]. With that in mind: [YOUR REQUEST].Tell it when to act.
It can think for minutes, which is great on hard problems, but on simple jobs tell it when to stop planning.
When you have enough information to act, act. If you’re weighing a choice, give a recommendation, not a survey of every option.Make it check its own progress.
Long runs used to fake status updates, so make it check every claim against the real result.
Before reporting progress, check each claim against an actual result. Only report work you can point to evidence for. If it isn’t verified, say so.Set boundaries.
Fable sometimes takes initiative you never asked for.
When I’m thinking out loud rather than asking for a change, give me your assessment and stop. Don’t make changes until I ask.Keep it running on long jobs.
Deep in a long session it can stop early, or panic about running out of context.
You’re working autonomously and have ample context. For reversible steps, proceed without asking. Don’t stop or suggest a new session.Give it a memory.
It works best writing lessons to a file and reading them back. A markdown file is enough.
Keep a notes file. Store one lesson per entry with a one-line summary of what worked and what didn’t. Update existing notes instead of duplicating.The rest are settings, not prompts. The one that matters is effort. It replaces model switching. Instead of dropping to a smaller model for cheap work, you turn the dial down.
This is Anthropic’s own benchmark. Fable 5 on medium beats Opus 4.8 on every setting, including max, and costs less doing it.
High is the default and the sweet spot.
The jump to xhigh buys five more points on the hardest problems. Max costs half as much again for barely two. Leave it on high, and push to xhigh when the job is genuinely hard.
Max is for the run you’d rather overpay for than redo.
How I actually build a loop
You don’t start with the loop. You start with the eval, a written test for the work so the loop can mark its own output instead of asking you.
Writing this very edition, I built one.
I write my newsletter in Claude Code, and a QA agent reviews every draft before I do.
First I needed the data.
I had Agent Browser, Vercel’s automation browser for AI agents, log into Substack and export all 110 of my past editions with their open rates and engagement.
Then Claude compared the winners to the flops and turned the difference into a scoring rubric.
Now a fresh agent scores every draft against it, cold. Fable plans, Sonnet executes, and the judge stays separate so nothing marks its own homework.
Under 95, it goes back with notes. This edition went from 51 to 95 before I read a word.
And it compounds. Every failure I catch by hand, I add to the rubric, and the loop never misses it again.
Build your own. Paste this:
Here are my three best [posts / emails / designs]: [PASTE THEM]
1. Study them and write a scoring rubric out of 100. Base the categories on what makes these good: [hook, structure, voice, clarity]
2. Draft a new one on [TOPIC]
3. Score the draft against the rubric as a strict reviewer
4. Under 95: rewrite and score again. Repeat until it passes
5. Show me only the final version and its scoreThat’s the whole loop in one chat. Build your rubric and yours does the same.
Your data is the moat
Every AI has read the same internet, so everyone’s output sounds the same. It hasn’t read your engagement stats, your best work, your flops.
Your own data is the moat. Feed it into the loop and the work is yours, not everyone’s.
The check is the job now
Start with one small loop this week. Claude Code has two commands for it:
/goalsets the finish line in one testable sentence/loopkeeps Claude working until that goal is true
The first loop feels like more effort than doing the task yourself. The tenth time, you’ve forgotten the task existed.
Writing the loop is the easy part. Writing the check is the job.
A prompt fails in front of you and you fix it on the spot. A loop fails while you’re asleep, handing you 50 finished pieces of work that are all subtly wrong.
The check you wrote is the only thing between you and that pile. So the human doesn’t leave the room. The human moves, and the job becomes:
Define “done” before anything runs
Read the checks, not the drafts
Keep your hand on the publish button
Slop is a failure of judgement, not tooling. Loops make that truer. They’ll produce more of whatever you tell them to accept.
Run this system yourself
I’ve just opened the alpha for the exact system in this issue.
The loops, the rubrics, the QA agent, set up to research, write and design your content (infographics and animated GIFs included).
I’m only letting 20 people in. Use code ALPHA at checkout for 50% off. First come, first served, and it’s the cheapest it will ever be.
Stay curious, stay human, and write the loop, not the prompt.
— Charlie
Ways we can work together
Become a Paid Subscriber: Get full access to every past issue of MarTech AI and support future ones.
Claude Content Engine: The Claude engine behind 350k+ followers (across socials) and 100 million impressions a year.
What's your scoring mechanism? Is it a prompt to the judge to score the text from 0 to 100 using the rubric and some framework following (like "hook, structure, voice, clarity"). I'm asking because I have something similar but without a loop. My writing agent has a hook to score it's work in the end taking my edits into account. But it doesn't work so well.