A migration that became a rebuild: Most of the week went on the new Fidero docs site. The plan was to move the old version over from its subdomain but instead I started from scratch: an information architecture spec first, then ~30 pages built in batches, all with Fable (Anthropic's new model). Amazing how much time it takes to do things like this properly. Worth it though: getting your champion excited is all for nothing if you don't stand up to scrutiny from their technical team. The marketing site sells the outcome, the docs make them believe it's legit. Docs

Fable to come off plans on Sunday: My approach – keep paying, but plan with it, don't execute with it. Strategy and specs get the expensive model, then a fresh session executes with Opus, unattended through to PR (pull request). A model doesn't have to be twice as good to justify twice the price: a wrong decision costs far more than the few extra dollars a Fable-written spec does.

While you're at it, pin your models: custom agents set to "inherit" will burn Fable usage on jobs that don't need it. Pin each one to the model the job warrants e.g. Opus is fine for code review. Skills take model and effort overrides in their frontmatter (the properties block at the top) too. One-time setup and you stop having to remember to switch.

The copy-and-paste monkey: Multi-model 'council' tools collect each model's answer in a blank envelope, so no model ever gets challenged or hears a rival take. What I actually want is two models in a room. The workaround for now: tell each one "you're in a dialogue with another AI, I'll relay the responses". Then I'm just the copy-and-paste monkey while they talk it through.

On my radar

Cowork is going mobile: Claude's desktop app is getting agents that keep running after you shut the laptop. The unresolved question is what happens to local file access and credentials once execution moves to the cloud. Announcement

Teams are losing engineers over AI: hearing of engineers walking (or being let go) over resistance to the direction. It's rarely about the tools. The attachment is to the craft of coding itself, and that's a harder conversation than any rollout plan.

Wikipedia's guide to spotting AI writing: editors have catalogued every tic and it's the best free checklist I've seen for stripping the AI tells out of your own drafts. Guide