The last day of a smarter model

So, here’s a situation that’s going to become normal and that nobody really has a playbook for yet: you get temporary access to a model that’s meaningfully smarter than your daily driver, and you know the exact day it goes away. Anthropic’s Fable 5 was included in subscriptions for six days. I came back from my off-grid holidays with about one day left.

The obvious move is to just use it, right? Point it at your backlog, watch it clear tickets faster than usual, feel good about it. That’s what I think a lot of people did with their window, and I think it’s the wrong call. Ticket-clearing is exactly the work the cheaper model already does fine. Spending frontier hours on it is like hiring a structural engineer and asking them to paint the fence.

The question I ended up ranking everything by: what can THIS model produce that next week’s model can’t? Not “what’s important”; importance is a trap here. My broken homelab alert delivery was important and urgent, and also completely fixable by the everyday model. It went to the backlog.

What made the cut was judgment. Decisions, specs, taste, the reasoning behind choices. So the plan became a spec factory: have the smart model make every hard decision it could reach, write each one down as an executable document, and leave the typing to its successor.

What a spec has to survive

A spec written by a smarter model for a dumber one has a documented failure mode, the instructions can just overwhelm the executor. So every spec followed one contract: fresh-session executable, all decisions pre-made with a one-line why, and every step ends with a mechanical acceptance check. An exact command and its expected output. Never “verify it works”.

Four specs came out of the first block: a health API extension for my status page, a deep-link plan for a WASM Bible reader, an Ansible migration for the homelab (a Pi 5, an old Pi 2, and a Hetzner box), and a DNS failover decision that had been nagging me for months.

That last one is a nice example of why the judgment layer matters. The textbook answer for DNS failover is keepalived, a floating IP between resolvers. The model rejected it for a reason I hadn’t articulated: my clients get DNS through Tailscale, and a LAN VIP would only be reachable through the subnet router, which runs on the exact machine whose failure it’s supposed to survive. The mechanism dies with its own trigger. Two lines in the Tailscale admin console do the job instead. That paragraph of reasoning is worth more than any amount of generated YAML.

The part where the specs turn out to be wrong

Writing the specs was maybe a third of the day. The rest was adversarial review, every spec went to a separate reviewer agent briefed to be hostile, and none of them survived the first round.

The reviewers caught things I find pretty humbling in hindsight. The DNS spec’s failover test required stopping a service over SSH with sudo, and my Pis run scoped sudoers where mutating commands need a password, so the command would just hang forever in a non-interactive session. The Ansible spec managed a cron job with ansible.builtin.cron, which quietly stamps marker comments into your crontab, guaranteeing the spec’s own “nothing changed” parity check could never pass. A rude default, honestly. The health API spec had a rollback procedure that rebuilt the broken code instead of the old code, because the deploy rsync had deleted the old source before anything got tagged.

Every one of those would have burned an evening of the successor model’s time, or worse, taken down DNS for the house.

Two to three review rounds each, until a clean approval. Then the closing move, and if you take one thing from this post, take this one: I handed a finished spec to the weaker model, cold, no context, and asked it to explain its execution plan back to me. Not to run it. Just to read it and say what it would do.

It read everything right. It also found seven small gaps, including a Go testing deadlock where the spec demanded failing tests before the types they referenced existed, so nothing could even compile. The format survived contact with its actual audience, which is the only review that counts.

Honest limits

The specs are validated as readable, not as executable, those are different claims. The smoke test says the successor understands the plan; whether the plan survives contact with live infrastructure is next week’s news. And the spend limit killed three reviewer agents mid-run at one point, which cost me an hour of resuming and re-briefing. Frontier windows and rate limits are an annoying combination, and I don’t have a fix for that.

There’s a bigger loop here I’ve only started on as well. The same day produced a handoff document where the smart model predicts, in writing, where the cheaper model will fumble on my setup specifically, and a rewrite of my two most judgment-heavy workflow files with those predictions baked in. Whether the predictions hold is perhaps the most interesting experiment of the lot, the model wrote down how its successor will fail, and starting tomorrow I get to watch it be right or wrong.

The backlog is queued, one spec per fresh session, cheapest model that can follow instructions. First up is the security findings doc, because the smart model’s parting gift was pointing out that a token file on my workstation had been world-readable since October.

I’ll write up what the successor actually does with all this once it’s had a week.

Update, 2026-07-15, nine days later. The successor is real and it shipped. The first thing through the queue wasn’t the security doc after all, it was the weekly review system, because the enforcement gap it closes is exactly what let this post slip its own publish date. That system has been running in production since the 13th, first report generated, two small bugs found and fixed after contact with reality, both visible in the commit log. The token file got its chmod today, rotation queued right behind it. The six-day window itself turned out to be soft as well, Anthropic has extended it twice since, both times on the supposed last day. The deadline was real when it shaped the day, it just refused to arrive. Which is its own small lesson, believing the hours were scarce is what forced the ranking question in the first place, I’m not sure the specs exist without it. So the week-later write-up promised above now has actual material, where the predictions held and where they didn’t is post #2.