If you’ve been anywhere near the AI corner of the internet lately, you know that Fable 5 access is quickly coming to an end. July 12. That’s the cutoff. After that, those of us in the research preview go back to the regular model lineup like everybody else.
Now, I’ve seen a lot of people spend their remaining access generating increasingly elaborate cat pictures (that’s a technical term meaning “not much”). I decided to try something different: use the window to do the work that actually requires a frontier model, and package everything else so that ordinary models can finish the job after the window closes.
Here’s how that went.
Where the idea came from
I’ve been building CooplyHQ, a platform for homeschool co-ops. Somewhere in a planning conversation, a simple observation surfaced: compliance with state homeschool law isn’t a co-op problem. It’s a household problem. Every homeschool family in a regulated state has to file notices, keep portfolios, and hit evaluation deadlines — whether they’re in a co-op or not. And most of them aren’t.
That meant the compliance tool I kept wanting to bolt onto CooplyHQ was actually a different business with a much bigger front door. So we vetted it the boring way: searched the market for name conflicts, ran the USPTO trademark database, checked what parents actually type into Google when they’re panicking about a filing deadline. The answer to that last one is pretty much the product name: HomeschoolBinder. Descriptive name, weak trademark, strong search intent. I’ll take that trade for an SEO-driven product. Registered the .com and the .app the same day.
The part that actually needed Fable
Here’s the thing I had to get honest with myself about. The app — the actual software — is ordinary CRUD. Forms, deadline math, PDF generation. Any competent model can build that in August. Spending my Fable window writing Laravel controllers would have been like hiring a master carpenter to carry lumber.
What genuinely needed the horsepower was the content and the judgment:
Fifty-one state law guides, cited to actual statutes. This is the growth engine — parents search “florida homeschool requirements,” and the page that answers correctly, with citations to Fla. Stat. § 1002.41 instead of vague hand-waving, wins. The catch is that getting statute law right is unforgiving work. Summarizer sites routinely get subsection letters wrong (I caught my own draft citing the wrong one — the statute’s own text is the only source that counts). And the law moves. In the course of two days of research, the agents discovered that Connecticut quietly passed a notification law six weeks ago, Iowa rewrote its homeschool statute in May, and New Hampshire has a bill sitting on the governor’s desk right now that would gut half of what its page says. Every “current” guide on the internet that was written from training data is already wrong somewhere.
My solution: one prototype page (Florida), built carefully by hand to lock the structure. Then a validation script — an actual program that checks every page for schema markup, citation format, FAQ integrity, dead links — and a fleet of parallel agents, each assigned a batch of states, each required to pull statute text from the legislature’s own site, and each unable to finish until the validator printed ALL CHECKS PASS.
Was it smooth? It was not. Agents occasionally decided to “wait for the other agents” (there were no other agents). Twelve parallel researchers rate-limited the web proxy into the ground. I hit my session token limit at what my clock said was a reasonable hour and what the server insisted was 2 AM. At one point I pressed the stop button because things looked hung and accidentally killed the whole fleet mid-flight. And here’s the lesson from all of that: because the work lived on disk and the quality bar lived in a script, none of it mattered. Every recovery pass just took inventory of what existed, what passed, and what was missing, and relaunched exactly the gap. Fifty-one pages, all validated, roughly two days.
The hand-off pack
That left the app. My first instinct was the same as everyone’s: write a big detailed PRD and hand it to the cheaper agents in August.
But that’s mostly wrong. A PRD describes intent, and every agent that reads it re-derives the design a little differently. What actually keeps a fleet of ordinary models on the rails is the same thing that kept the content generation on the rails: artifacts that constrain, not prose that advises.
So the pack the August agents will inherit looks like this:
- A JSON schema for state compliance rules — seven deadline-math patterns, a closed vocabulary. If a state doesn’t fit, that’s an escalation to a human, not an improvisation.
- Florida fully encoded as the reference module, straight from the verified research.
- Fifteen golden test cases covering the nasty edges (a filing window landing on Feb 29, anniversary dates anchored to Feb 29, DC’s “15 business days” counted across Labor Day) — every expected date machine-verified by a small Python oracle before any production code exists. The future PHP engine has to agree with the golden file, byte for byte, or it isn’t done.
- A database schema, an engine spec, nine architecture decisions marked “do not relitigate,” and a build plan of twelve slices — each sized for one agent session, each with a pass/fail gate.
- And an operator’s guide for the human (me): which model tier per slice, what to eyeball, what smells like trouble. Only one slice — the deadline engine itself — is flagged “strongest model available, no exceptions.” Everything else is deliberately boring enough for anything.
The PRD is in there too. It’s the thinnest document in the folder. That’s on purpose.
So — what will you do?
The window closes July 12. If you’ve got remaining access, here’s my unsolicited advice: don’t spend it on work an ordinary model can do next month. Spend it on the judgment — the research that’s unforgiving, the schema that everything else hangs on, the test cases that define “correct.” Then write those decisions down in a form that can’t be argued with.
Remember, the best use of a powerful tool isn’t always the most impressive demo. Sometimes, even in the AI era, the KISS principle is best: do the hard thinking once, freeze it in executable form, and let the cheap tools carry the lumber.
What will you do with yours?