A few weeks ago I wrote that the bottleneck in software had quietly moved. It went from can the machine build it, which it demonstrably can, to can you describe what you want clearly enough that a machine can run with it for hours without you. I promised a follow-up. Not tutorials, actual products, with the prompts, the costs, and the failures. Including the embarrassing ones.
This is that follow-up. And there's a clock on it.
When Fable 5 came back on July 1, Anthropic included it free on paid plans, first through July 7, then through July 12 after the pushback. So the free window closes on July 12, and after that it's pay-per-use credits. For the last week and a half I did the least sophisticated thing imaginable with a frontier tool. I tried to build as much real stuff as I possibly could before that meter starts. Not to hoard it, but to find out, on my own money and my own name, where the new bottleneck actually bites.
Here's what came out. Three apps, a content system, and two reusable skills. Some shipped. Some are honestly mid-build, and I'll tell you which is which, because the half-built ones are where the real lessons are.
One thing up front, because it's the whole game. The difference was never the model. Same Fable 5 everyone with a plan can open. The difference is the inputs. My sources, my voice, my taste, and a clear enough brief that an agent could work unattended. Generic in, generic out. That part doesn't expire on July 12.
1. The system that makes my content
The first thing I built is the one I needed most. A content engine. One chain, research in, then idea, script, video, publish, designed to collapse a two-day-per-video workflow into an approval and a coffee.
The shape of it, briefly: research accumulates on its own across the topics I care about, roughly 2,000 sources so far, a daily brief proposes three ideas already in my voice, I pick one, and a finished script comes back before the coffee goes cold. I've written that pipeline up stage by stage, including the stage I refuse to hand over, in the content engine post, so I won't repeat it here.
What belongs in this ledger is the seam. Every editorial decision in that chain is still mine, and the sprint didn't change my mind about which ones those are.
The most useful thing I learned building it was a failure. My first automated video render looked stiff and uncanny, and I nearly wrote the conclusion "AI avatars aren't there yet." Then I rendered the same script by hand, a proper video avatar, real B-roll, captions, and it looked genuinely good. The blocker was never the avatar. It was the editorial cut I'd tried to skip. The part you can't hand to the machine is the judgment about what's good, not the generation itself. The full post-mortem, frame by frame, is in The AI Re-Render Came Back Worse.
There was a second failure worth your time. An early version tried to pull discussion from a platform that blocks scrapers. My clever fix, driving a logged-in browser around the block, got my home IP rate-limited and an account banned inside a day. I ripped it out entirely. The tool now reads public search results only, and if it can't reach something it says so. A tool that admits "I couldn't reach this" beats one that invents plausible data. Write that on the wall.
Status: the brief, the reply-to-develop loop, and the scripting all work and run daily. The last mile of video polish I've pulled back to manual on purpose, until I decide how much of the taste is safe to automate. Building in the open means telling you that instead of pretending it's a closed loop.
2. The strange little internet experiments
The second promise was a portfolio of small, strange experiments. A test of one question. Can a well-specified, AI-built product find an audience with zero marketing budget? You can't answer that without shipping some.
A studio portfolio with a moving heart. I took an awards-tier agency site apart to understand why it wins, then rebuilt its signature move cleanly in my own stack. A homepage where a globe of work expands into an infinite scrolling gallery. Most developers can't ship that interaction. Fable 5 and a fleet of subagents did it in days. The showcase pieces are placeholders behind a no-index wall while I swap in my own work, so what I'm showing here is the engine I built, not the demo content sitting in it. The extraction method, and why that site wins, is in Rebuilding an Awards-Tier Interaction With AI Subagents.
A viral-app skeleton, and its first launch. The more interesting build underneath the fun one. A reusable skeleton for launching "pay-for-status" micro-apps as config and a skin, rather than a new codebase each time. The first launch riding on it is a king-of-the-hill game. One seat, and whoever's paid the most holds it until someone outbids them. It sounds like a toy. Under the hood it's 398 passing tests, row-level database security, per-IP rate limiting, sanctioned-region geo-blocking, and an atomic takeover that can't be raced or underpaid. The visual went through three designs I rejected before landing on the one I liked, a black monolith with a molten-gold seam that charges as you type your name. Why a silly app got production infrastructure is its own argument, made in I Built Production Infrastructure for a Silly App.
Full disclosure. This one runs on test-mode payments right now. It is built, not launched. Real keys, legal pages, and a data wipe stand between it and its first real dollar. When it takes real money and I have real numbers, that's its own post. No pretending here.
3. The one I've needed my whole life
The third thing I teased as "something bigger, for a problem I've struggled with my whole career." Here it is. An ADHD focus tool called Anchor. It catches the moment your attention drifts off-task and walks you back. No task lists, no streaks, no red, no shame. A browser extension watches only which site you're on, entirely on your own machine. Nothing leaves it.
I built the working thing in about a day. But the part I'm proudest of has nothing to do with the app itself. It's the decision about what to build first underneath it. The obvious move is accounts and billing. Instead the first backend milestone is an opt-in, anonymous evidence pipeline that measures whether the app actually helps, and publishes the aggregate, including results that come back inconclusive. There are over a hundred ADHD focus apps and almost no published evidence any of them work. In that vacuum, honesty is the product. The longer version, including why I won't charge for it until it can prove it works, is in I Built the Focus App I Actually Needed.
Status: it runs and it's deployed, but the extension hasn't been hardened in the wild yet and it isn't in the store. Mid-build, on purpose, in public.
The receipts, and the ones I still owe
I promised numbers, good or bad, so here's the honest ledger of a week and a half.
- Three applications, a content system, and two reusable skills, built or substantially built by one person.
- Time: the video workflow the content engine targets used to cost me two days per video. The studio portfolio's moving front-end is the kind of build I'd expect to wait two to three weeks for from a specialist front-end shop. Phase one shipped in a day.
- Cost: almost all of it ran on my existing Claude subscription, inside the free Fable window. The only real cash on top was roughly $200 in HeyGen credits. Both of those are my own estimates rather than invoiced figures, and I'll show the per-project breakdown in the deep-dives.
- What isn't done: the content engine's last mile is manual, the king-of-hill app is on test keys, and Anchor isn't in the store. Building in public means the "not done" column is part of the report, not hidden from it.
I'm also holding one back deliberately. A client-grade studio website I shipped live this week for a real service business. Proof that you can move this fast on client work without it ever looking machine-made. That one gets its own write-up so I can do the craft decisions justice.
What I'm keeping after July 12
Here's the part that matters even if you never touch Fable 5, or if the free window shuts before you do.
The sprint didn't just produce apps. It produced methods, and I packaged two of them as reusable skills. One is an operating mode. I act as the planner and architect, and delegate the heavy building to cheaper, faster models running in parallel, verifying their work instead of trusting it. The other maps audience demand across every major platform before I make anything, so the ideas come from real signal instead of a guess. There's a third in the wings, the site-extraction method behind that portfolio rebuild.
None of them depend on Fable specifically. The orchestrator is whatever model you launch. Fable made it faster and cheaper, but the method runs on any capable model. When the free window closes, the skills don't. That's the actual asset from a fortnight like this. Not the token allowance, but a repeatable way of working that outlives the model that taught it to you.
The bottleneck moved. It isn't the code anymore. It's the brief. And a good brief is a thinking skill, not a programming one. If you've ever written a clear scope of work, you're closer to shipping software than you think.
Frequently Asked Questions
What is Claude Fable 5?
Fable 5 is Anthropic's top-tier model. It arrived in early June 2026, was pulled on June 12 when US export rules landed on it, and came back on July 1. Anthropic included it free on paid plans for up to half of your weekly usage, first through July 7 and then through July 12 after the pushback. After that it moves to pay-per-use credits.
What does Fable 5 cost after the free window?
Ten dollars per million input tokens and fifty per million output, which is double the price of Opus 4.8. Same tool, very different meter, which is why the fortnight before the window closed was worth spending on real builds rather than tutorials.
Can one person really build three apps in a fortnight with AI?
Built, yes. Finished, no, and the distinction matters. Three applications, a content system, and two reusable skills came out of the sprint, but the content engine's last mile is still manual, the king-of-the-hill app runs on test payment keys, and the focus app is not in the Chrome store. The honest ledger includes the "not done" column.
What actually changed about building software?
The bottleneck moved from the code to the brief. Machines can demonstrably build the thing. The hard part is describing what you want clearly enough that an agent can run for hours without you, which is a thinking skill rather than a programming one. Anyone who writes a clear scope of work is closer to shipping software than they think.
Was any of this dependent on Fable 5 specifically?
No, and that is the point of packaging the methods as skills. The orchestrator is whatever model you launch. Fable made the sprint faster and cheaper, but the way of working, a planner delegating to cheaper models in parallel and verifying their output, runs on any capable model and outlives the window.
Want help building systems like this for your organization?
See what a system like this could do for your workSo now I'll turn it around, because learning out loud only works if it's a conversation.
What did you build this week? Not what you read about. What you actually made, shipped, or broke and fixed. Whether you've been sprinting before the window closes or just circling the idea, tell me what you put into the world. I read every reply, and the best ones end up in the next post.
The window's closing on one thing. It's wide open on the thing that actually matters.
