Real quick before I start, you'll notice this entire issue is free.
This is a change we are doing to the Monday versions only of Build Notes. Mondays will always be some kind of deep dive and the full issue is free. We are PACKING Mondays with value. While Thursdays will be our usual style, free + paid tutorial.
We want to make sure that Build Notes stays highly valuable to every reader. Also, we are adding new bonuses to Build Notes+ members. Details on that coming soon.
Now here's issue #016...
Tomorrow night at 11:59pm Pacific, the smartest AI ever handed to the public goes back behind a paywall.
You've got about 36 hours.
This issue is your battle plan.
Quick recap for anyone who missed the soap opera... Anthropic released Claude Fable 5 on June 9th. It's the first public "Mythos-class" model, which is their way of saying "the stuff we used to keep locked in the basement." It benchmarks more than 10% above Opus 4.8, and it was built for work that runs for hours or days without you babysitting it.
Then the government got involved, it went offline, and Anthropic brought it back on July 1st for one week inside paid plans (up to 50% of your weekly usage).
That week ends tomorrow, July 7th.
After that, Fable 5 moves to usage credits: $10 per million input tokens and $50 per million output tokens. Anthropic says they aim to put it back into subscriptions when they have the capacity, but there's no promised date.
Remember Thursday's Profit Dashboard issue? Fable 5 wrote that email. Now you know why I made you wait until Monday for this guide.
Four things I'm teaching you today:
- How to prompt Fable 5 (most people are doing it wrong)
- The image trick that cuts your token bill 60%
- How to run marketing loops on it
- The 24hr play that keeps Fable's brain in your business after it leaves
Prompting Fable 5 Means Unlearning
For two years I've taught you to write detailed structured prompts. ROLE, TASK, numbered instructions. QUALITY IN, QUALITY OUT.
Fable 5 flips part of that on its head.
Anthropic's own prompting guide says prompts written for older models are often TOO prescriptive for this one. Hand it a 40-step checklist and it will follow your checklist right past a better solution it could see and you couldn't.
What it wants instead is the goal and the WHY. Here's the exact template from Anthropic's guide:
"I'm working on [the larger goal] for [who it's for]. They need [what the output enables]. With that in mind: [request]."
Here's what that looks like in marketer-ese:
"I'm rebuilding the welcome sequence for my $249/yr newsletter. Readers who finish the first three emails renew at triple the rate of readers who don't. With that in mind: rewrite emails 1-3 so more people finish them."
That's the whole prompt. No 14 rules about tone. Fable connects dots you didn't draw.
Second thing to know: the effort dial. Fable has settings from low to max, and effort is the lever that trades speed and cost for intelligence. Run high (the default) for real work and xhigh for the make-or-break stuff, like the offer page for your biggest launch of the year.
Third, and this is my favorite part of everything I read this week...
An Anthropic engineer named Thariq published a field guide built on one old idea: the map is not the territory. Your prompt is the map. Your business is the territory. The gap between them is full of "unknowns," and with a model this capable, YOUR ability to surface those unknowns is the bottleneck. Not the model.
Three of his moves, translated for marketers:
- The blindspot pass. "I'm about to run YouTube ads for the first time. I've only ever bought Google search. Do a blindspot pass: find my unknown unknowns and help me write a better prompt for you." It will tell you the questions you didn't know to ask.
- The reverse interview. "Interview me one question at a time about this offer. Prioritize questions where my answer would change the funnel." Flip the desk. Let the model dig the brief out of your head.
- Reference materials beat descriptions. Don't write 500 words describing your best ad. Paste the ad. "Make more like this, here's why it worked" beats every adjective you own.
Pay Per Pixel, Not Per Token
This one's been ripping around X all week, and it matters double once Fable goes to per-token pricing tomorrow.
A developer named Robert Youssef stopped sending his context as text. He sends it as an image.
Why? Because you pay for an image by its pixel size, not by the text packed inside it. Dense text crams about 3 characters into every image token, while plain text gets about 1. His system prompt, tool docs, and old history now go in as one PNG, and Fable 5 reads it straight back.
Same brief, roughly a third of the tokens. See why the ad buyer in me perked up?
There's already an open-source tool called pxpipe that does the conversion automatically, and in one demo a $42.21 Fable session dropped to $6.06.
Two warnings before you go screenshot-crazy:
- Exact strings stay as TEXT. Coupon codes, API keys, tracking IDs. In recall tests Fable read imaged text almost perfectly (13 of 15 on exact character strings) while Opus 4.8 got ZERO of 15. This is a Fable trick, not an every-model trick.
- It adds a little wait time while your machine encodes the image.
The marketing translation: your brand guide, your avatar research, and your swipe file are the bulky part of every prompt you send. Turn them into PNGs once and attach them forever.
Loops: What Fable Was Actually Built For
Back in issue #016's grandfather, issue #012, I taught loop engineering. Short refresher: a prompt is an instruction, but a loop is a finish line. You define what done looks like, the scope it can touch, and the stop rules. Then the agent works until it crosses the line.
Boris Cherny, the guy who built Claude Code, put it this way: "I have loops running that prompt Claude. My job is just to write loops."
When I wrote issue #012, loops were the future. Fable 5 made them the present, because it's the first model that holds a multi-hour loop without wandering off into the woods. It verifies its own output, dispatches little subagent helpers in parallel, and keeps notes as it works.
Three marketing loops worth running before tomorrow night, while the meter isn't running:
- The competitor teardown loop. Screenshot 50+ competitor ads (they're CHEAP as images now, see above). Loop until it delivers a messaging map, the patterns everyone's using, and 10 angles nobody in your market has tested.
- The funnel audit loop. Point it at every page in your funnel. Loop until each page is scored on headline, proof, and offer, with a prioritized fix list and rewrites for the three worst offenders.
- The content refresh loop. Feed it your old blog posts. Loop until every one passes an AEO checklist: direct answer in the first 50 words, question-style H2s, quotable standalone blocks.
And ALWAYS cap your loops. Max passes plus a dollar budget. A loop without stop rules is a leaf blower with the trigger taped down.
The 24hr Play: Burn It Into Skills
Fable leaves tomorrow. Skills don't.
A skill is just a markdown file of instructions your AI reads before it works (SKILL.md in the Claude world), and your CLAUDE.md is the same idea: a briefing your agent reads at the start of every run. They cost almost nothing and they never expire.
Fable 5 writes better skills than any model I've used. More precise, more complete, and ruthless about edge cases. And a skill it writes today keeps working on Wednesday, because Opus 4.8 and Sonnet 5 read the same files.
One tester had Fable rewrite his skills, then blind-graded the results running on Opus 4.8. The Fable-written versions won 12 of 14 matchups.
So today, while it's still included in your plan, run this prompt:
"Review my past sessions and my prompts. Cluster my repeated instructions and my repeated corrections. Write skill files that encode everything you'd want a less capable model to know about how I work. Then rewrite my CLAUDE.md to be shorter and sharper."
Then have it write one skill per workflow that makes you money. Your ad teardown checklist. Your email voice. Your offer math. Whatever you find yourself re-explaining every week.
Want to see what a professional-grade skill looks like before you start? Go peek at taste-skill, an open-source skill that stops AI from designing those generic purple-gradient websites. Locks, bans, pre-flight checklists. That's the level of detail Fable will write about YOUR business if you ask it to.
After tomorrow, the math changes. Sonnet 5 is sitting at $2/$10 intro pricing through August, so it becomes your daily driver, and Fable becomes the $1,000-a-day consultant you fly in for the jobs that move real revenue. Which is fine. That's exactly how I used to treat my best contractors.
But right now, for about 36 more hours, the consultant is sitting in your office for free.
Don't spend that time asking it for Instagram captions. Make the genius write down everything it knows before it leaves town... and then it won't matter which model answers the phone on Wednesday.
Continue with Build Notes+
This Monday deep dive is free in full. For complete Thursday builds and supporting files, join Build Notes+.
