Deep Dive · XiaoHu Explains

A Practical Guide to Claude Fable 5: 8 "Finding Unknowns" Prompt Patterns, from Blind Spot Scans to Post-Task Quizzes

Claude Code team member Thariq Shihipar's conference talk transcript: the bottleneck for new models isn't the model anymore — it's whether you can articulate your own unknowns
Quick Take
  • Anthropic Claude Code team member Thariq Shihipar gave a talk titled "A Field Guide to Fable" at the AI Engineer World's Fair conference, with a companion official blog post published the same day — reportedly reaching about two million reads within three days.
  • Core methodology: split "what you know" into four quadrants — known knowns, known unknowns, unknown knowns, and unknown unknowns. Reducing and pre-planning for these "unknowns" is the core skill for using the new generation of models well.
  • The talk offers 8 ready-to-copy prompt patterns, organized by before, during, and after doing the work, each paired with a ready-made bilingual example prompt.
  • A first-hand, counterintuitive disclosure: Claude Code's system prompt was recently cut by 80%. For the new model, examples actually constrain it — the direction has shifted from "giving constraints" to "giving context."
  • Thariq used this exact method to edit Fable's launch video from scratch, in an area — video editing — completely new to him.
Stance note: this is a talk by Anthropic (the Claude Code team) at the AI Engineer conference, paired with an official blog post of the same name, about how to use their own Claude Fable 5 well — it's vendor content. The talk contains almost no benchmarks, architecture details, or hard numbers for Fable; the bulk is methodology and first-hand observations on internal prompt engineering.
1What This Talk Is About

What Exactly Is This Talk About

Anthropic Claude Code team member Thariq Shihipar gave a keynote talk at the AI Engineer World's Fair conference, and simultaneously published a companion official blog post of the same name, "A field guide to Claude Fable 5: Finding your unknowns."

This talk is about something very specific: how far this generation of models — Fable — can go, the bottleneck has shifted away from the model itself and onto you, onto whether you can clearly articulate "what you don't know." The talk contains almost no hard metrics or architecture details for Fable — it's entirely about "how to collaborate with a stronger model."
📌
Why it's worth watching: Claude Code's system prompt was recently cut by 80%, and the direction has shifted from "giving constraints" to "giving context" — a first-hand look at a shift in Anthropic's internal prompt engineering. It also distills "how to use a stronger model well" into 8 copy-paste-ready prompt patterns, each with a ready-made example.

One word keeps coming up in the talk: unhobbling. The idea is that what actually constrains the model is often the framework you yourself impose on it, and your old prompting habits — frameworks that are essentially a reflection of "how well you understand the model." Change them, and you release the capability that was already there.

An analogy

Taking the unnecessary reins and blinders off a horse that could already run.

~2 million
Reported reads of the companion official blog post within three days of publication
80%
Recent reduction in Claude Code's system prompt, shifting from "giving constraints" to "giving context"
8
Number of copy-paste-ready prompt patterns distilled in the talk
Full talk · Chinese-English bilingual subtitles (~19 minutes, click cover to load) · Original video on YouTube (English) ↗
2Background · Prompts Are Shrinking

Models Get Smarter in "Spikes" — Which Is Why Prompts Should Actually Shrink

To understand why this generation of models means "fewer examples, more context," look at two pieces of background first. This section is setup — the main event is later.

The first piece of background is capability overhang: models don't get stronger evenly across the board — they suddenly jump way ahead on certain specific tasks. There's a viral post online asking "why can't a large model say which Pokémon names end in AW" — a regular chat model can't answer, even though it obviously knows every Pokémon's name. But Claude Code can answer, because it will pull down every Pokémon and write a script to filter for names ending in AW. The same "knowledge," but whether or not you give it a code execution tool makes a world of difference.

An analogy

Like ore veins buried at different depths — the tool is the pickaxe that determines which vein you can actually reach. Half the work of using a new model well is figuring out what new possibilities now exist.

The second piece of background is more counterintuitive: system prompts are getting smaller and smaller. Thariq breaks prompt engineering best practices into three eras, and you can see the trend.

The Sonnet 3.5 New Era
Small Prompt · Few Tools
Mostly relied on piling on lots of examples, teaching it to copy along.
After Models Got Stronger
Large Prompt · Many Tools
Could fit in more information and instructions — it started being able to just follow them, but examples kept piling up too.
The Fable Generation
System Prompt Cut by 80%
Examples actually constrain it now — it's more imaginative than the examples you give it. Give it context, not constraints — try not to write "don't do this."