An Anthropic Engineer's 'Field Guide to Fable': Closing the Gap Between the Map and the Territory

Thariq Shihipar, who works on Claude Code at Anthropic, took the stage at the AI Engineer conference to talk about his strong affection for Fable. He places it among Anthropic models people just remember, like Sonnet 3.5 new or Opus 4, and describes the experience of using Fable as finishing an RPG's tutorial and finally reaching the point where the open world unlocks. There's suddenly so much you can do and explore -- which is exciting, but also a little intimidating.

One thing he's learned working with Fable is that it's not just Claude that needs to be 'unhobbled' -- you do too. The plan, prompt, and spec in your head is just the map; the actual codebase and real-world constraints are the territory. Whenever Claude runs into something in the territory that isn't on the map, he calls that an 'unknown.'

Because Fable can explore such a large space, if you don't find your unknowns ahead of time, it's going to run into a lot of them. He proposes organizing this as a matrix -- known knowns (what you put in your prompt) versus known unknowns (things you know you haven't figured out) -- arguing that your ability to use Fable well ultimately comes down to how accurately you can match the map to the territory and surface your unknowns.

He reveals he built the very slide deck for this talk in about 4 hours the night before, using Fable -- a concrete example of combining speed with quality he was genuinely happy with. He tells AI engineers the world is looking to them to prove AI isn't just a fad, that it can genuinely make people more productive and save time -- and shares his own resolution for the year: be more productive, but work less, and spend more time with the people he cares about.

Building itself has gotten easier thanks to AI, he notes, but generating actual value remains hard and takes a lot of swings to find what's genuinely valuable. He closes with: go explore, make it real, and be less reasonable.