On Fable's launch day, an Anthropic member on why what contains the models is us
Field Guide to Fable — Thariq Shihipar, Anthropic · Thariq Shihipar
19 min total·Actually worth watching closely: ~7 min·2 must-watch clips
- 0:12 – 2:00Listen
The models are grown, not designed
The opening swaps the mental model for understanding models from physics to biology: there is no set of rules to derive from, only experience, observation and intuition built up slowly. The speaker is candid that the Anthropic team is also learning with the model as they use it.
Don't expect a manual. Judgment about the model comes from repeated hands-on use, the same way biological intuition does.
The speaker is talking through an argument the whole way and there's nothing on screen you need to watch — headphones on the commute loses you nothing.▶ Jump to 0:12Speaker · Thariq Shihipar - 2:00 – 4:40Listen
Claude Code removed 80% of its system prompt
Prompt engineering is turning over a generation: newer models want a smaller prompt, examples end up boxing in a model that is more imaginative than the examples, and a stack of "don't do X" rules ties your own hands. The principle becomes give context, not constraints.
The examples and prohibitions you carefully wrote may be exactly why the model isn't performing the way you expected.
Argument-driven, carried by sentences rather than slides; even at the moment the prompt cut comes up, no before/after prompt goes on screen. Listening is more efficient than watching.▶ Jump to 2:00Speaker · Thariq Shihipar - 4:40 – 7:13Listen
What really contains the model is us
The core of unhobbling: give the model tools and capability jumps in spiky ways — the same Pokémon question plain chat can't answer gets solved once Claude Code writes a script. Rather than stretching the context window without end, hand it arms like the Bash tool and let it search and build its own context.
The capability gap is often not in the model but in never having given it a chance to act.
The Pokémon example is told, not run live, so there's no need to sit on the screen — but this stretch is worth rewinding for a second listen.▶ Jump to 4:40Speaker · Thariq Shihipar - 7:13 – 9:20Listen
The map is not the territory
The prompt and spec are the map in your head; the codebase and the real constraints are the actual territory. Wherever the two don't line up is a decision point you never thought through and never wrote down. Fable traverses such a large area that the ceiling ends up sitting on whether you can find your own unknowns.
The bottleneck on output quality has moved from the model's capability to your ability to map your own blind spots.
The theoretical hinge of the whole talk, built up layer by layer by the speaker. The concept isn't complicated but you have to follow along — listen with attention rather than coding through it.▶ Jump to 7:13Speaker · Thariq Shihipar - 9:20 – 14:15Listen
Four techniques for four kinds of unknown
Blind spot pass — have Claude go through the codebase, Git and Slack to surface the problems you never noticed. Brainstorms and prototypes — ask for four wildly different directions at once and let your own reaction set the course. Reverse interview — have Claude interview you and force out the decisions you never specified. Reference as map — hand over existing code that shows the target shape instead of writing a spec by hand. Along the way, how interaction itself changed across generations: Ask User Question could barely be called on Opus 4, could ask forty questions about a spec by Opus 4.5, and with Fable produces a whole HTML report with the questions embedded inside it.
Each method treats a different kind of unknown; the reverse interview is the most counterintuitive and the most effective — the model asks you, not the other way around.
The four techniques are delivered one after another with no live demo and no screenshots, so take notes as you go. Open a page, write the four names down first, then listen for the details.▶ Jump to 9:20Speaker · Thariq Shihipar - 14:15 – 17:00Listen
Good, fast, cheap — pick three; be unreasonable
The economics changed and the implicit trade-offs changed with them: what used to be pick two can now be all of it. So don't rush to prioritize and do one thing — do all of it, and force reality to show you the real trade-off. The best way to do more ambitious work is to reframe your own ambition.
The two options you'd have cut under the old rules very likely don't need cutting at all now.
The most charged stretch of the talk, and an argument rather than a display of material — the rhythm in the voice matters far more than the screen.▶ Jump to 14:15Speaker · Thariq Shihipar - 17:00 – 19:07Listen
Building got easier; generating value didn't
The close lands a reminder: AI engineers are the ones most likely to get absorbed in building itself and in tuning their setups, but finding what is actually valuable still takes a lot of swings. Building was never the point.
The smoother the tools get, the more you have to watch for mistaking "this feels great to build" for "this is the right thing to build."
A closing passage that slows down and leaves something for the audience to sit with. Nothing to look at, but worth staying to the end for.▶ Jump to 17:00Speaker · Thariq Shihipar