← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Tue Jul 07 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Field Guide to Fable (Thariq Shihipar, Anthropic)

fableclaude-codeanthropicai-engineercapability-overhangunhobblingunknownsmythos-classprompting

Field Guide to Fable (Thariq Shihipar, Anthropic)

Thariq Shihipar (Anthropic member of technical staff on Claude Code) at AI Engineer World's Fair 2026 (~19 min). Recorded within the week Fable rolled out to the public (Shihipar: "Fable is back… we're rolling it out later today"). The talk's function is to teach people how to work with the new Mythos-class tier.

The four parts

  1. Unhobbling Claude — how you understand and unlock Claude
  2. Finding your unknowns — matching the map to the territory
  3. Dealing with the grief — the transition-cost side of Fable-class work
  4. Being unreasonable — the "tradeoffs are not real" reframing

Part 1 — Unhobbling Claude

Central metaphor: the map is opening up. Fable is where the tutorial ends and the open-world RPG starts.

Capability overhang

Explicitly named. Models get smarter in spiky ways — new capabilities emerge unevenly, and the tools you give the model decide which spikes you can reach. Shihipar's canonical example:

Ask a chat model "which Pokémon names end in AW?" → it fails, even though it knows every Pokémon by heart. Ask Claude Code → it fetches the list and filters for AW in seconds (Croconaw + Drednaw). Same underlying knowledge; a code-execution tool converted it into a solvable task.

See Capability Overhang for the concept page.

Three prior capability jumps (Shihipar's timeline)

  • Chat models needed context to be pasted in. Insight: give the model arms (bash tool, environment access) and it builds its own context. → Claude Code.
  • Claude Code had to be prompted. Insight: give the model the ability to wake itself and work proactively → Claude Tag (see Claude Code).
  • System prompts kept growing. Latest reversal: Claude Code removed 80% of its system prompt for the new model class. Examples now constrain the model — it's more imaginative than the examples given. Give it context not constraints; avoid "do not do this" instructions that were necessary for older models.

Product-level Fable-class progression that vault should track

  • Ask-user-a-question tool (Shihipar built it). Opus 4: barely worked, had to tweak the tool to get any calls. Opus 4.5: could interview the user with 40+ questions on a spec. Opus 4.8 / Fable: builds full HTML reports with questions embedded. Not incremental — an interaction paradigm change.
  • Markdown → HTML output. Initially Markdown was the good rich output; then plan-mode Markdown became a human-facing artifact; now with Fable, Claude builds in-depth HTML reports. A visible spike of the capability overhang.

"Biology, not physics"

Shihipar's frame: model behaviour is empirical / organic, not axiomatic. Cites Anthropic's Biology of a Large Language Model as the paper worth reading. Practical implication: build intuition, don't just look up rules — the rules haven't been written.

Part 2 — Finding your unknowns (the "map is not the territory")

The Alfred Korzybski line as a working discipline. Shihipar's unknowns matrix:

Known Unknown
Known known knowns — usually what you write in your prompt known unknowns — things you know you haven't figured out
Unknown unknown knowns — so obvious you wouldn't write it down, but you know it when you see it unknown unknowns — haven't considered at all

Fable is the first model where he felt he had to explicitly find his unknowns — because Fable traverses so large an area that unspecified decision points become the dominant risk. The prompt is bottlenecked by your ability to match map and territory.

Six practices for finding unknowns (Shihipar's operational payload)

  1. Blind-spot pass. "I'm working on a new auth provider I know nothing about — can you do a blind-spot pass and help me figure out my relevant unknown unknowns?" Also works for learning new fields (Shihipar used it for color grading in video editing).
  2. Brainstorms + prototypes. For design specifically — surfaces unknown knowns. "Make me an HTML page with four wildly different design decisions so I can react." Know-it-when-you-see-it.
  3. Interviews. Ask Claude to interview you. Add context — "prioritize questions that would change the architecture." Direct heir to Matt Pocock's Grill Me skill.
  4. References. "Here's some code that represents what I want to be done — read it, understand it, use it to start your work." Passing another map. Also works with HTML mockups for React components.
  5. Implementation notes. While Fable runs, ask it to log unknowns it hits. Then review deviations post-run to understand where the map and territory diverged.
  6. Quiz me. After the run, have Fable quiz you on what happened. Keeps you in the loop; lets you represent the work in the PR / merge.

Part 3 — Dealing with the grief

Shihipar's most personal move. First time with a Mythos-class model, he felt "a huge sense of gain, but also a sense of loss." He returned to his old YC startup codebase (30 people, constantly forced into tradeoffs) — and found "the things that would have taken me weeks I could do in hours."

"How can you not laugh? Also, how can you not cry? … I really, really loved programming and writing code by hand … but I also remember just staying up late nights trying to debug, working on things for weeks without working. I just remember swimming in failure … as much as I enjoyed those highs, I can't go back."

Frame: the only way out is through. Staying in the loop, unhobbling the model, and coming out with much more on the other side.

Adjacent to Justin Sung's Higher-Order Thinking framing of "you and AI are not in the same race" — but from the practitioner-mourning side rather than the career-strategy side.

Part 4 — Being unreasonable

Anthropic-cultural frame Shihipar wants people to leave with: tradeoffs are not real.

  • The reasonable move: write priorities, trade off between them.
  • The Anthropic move: what if you just did all of it, and forced reality to show you the tradeoff?
  • The classic "good, fast, cheap: pick two" → "the math of Claude and Fable really changes how you think about tradeoffs — now it's pick three."

"The best way to do more ambitious work is to reframe and make ourselves more ambitious. Because the only way to prove that agents work is to do the best work of our lives faster than ever before."

Concrete demonstration: Shihipar built the entire keynote deck the night before, in ~4 hours, with Fable.

Closing frame: "Building is easier, but generating value is still hard." AI-engineer trap = spending on process instead of value.

2026-07-24 sequel — Anthropic official blog

Shihipar published The New Rules of Context Engineering for Claude 5 (Thariq Shihipar) on the Anthropic blog on July 24 2026. That piece is the operational "how" to this talk's "why":

  • Same 80%-reduction claim, now with the qualifier "with no measurable loss on our coding evaluations."
  • Six named pattern shifts as an authored rubric (rules→judgment, examples→interfaces, upfront→progressive-disclosure, repetition→simple-descriptions, manual-memory→auto-memory, simple-specs→rich-references) — the WF talk had the intuitions; the blog freezes them as first-party guidance.
  • Anthropic's claude doctor / /doctor command as the tooling for the ongoing rightsizing pass — first-party version of Pocock's deletion test.
  • Adds artefact-specific rules for system prompts vs CLAUDE.md vs skills vs references — the blog is more prescriptive than the talk, which stayed at intuition level.

See Context Rightsizing for the vault's concept page synthesising both artefacts into a discipline distinct from context engineering.

Load-bearing convergences with the vault

  • Design Concept (Brooks) — Shihipar and Pocock, one conference talk apart, both arrive at the same prescription: interview the human before the agent starts fanning out. Now a two-source concept.
  • Higher-Order Thinking — Shihipar's unknowns matrix operates purely at the higher-order layer (relationships between decisions, contextual mapping). Third 2026 source pointing at the same skill.
  • Claude Code — 80% system-prompt reduction; ask-user tool progression; Claude Tag as proactive-work capability. Direct first-party updates.
  • Mythos-Class Models — capability-overhang mechanism named for the class.
  • Skills (Claude Code) — the smaller-system-prompt, less-examples, more-context discipline is the vault's Skill Checklist recasted for the model side.
  • Task Imagination (Nate B. Jones) — Shihipar's "task imagination" is what "being unreasonable" operationalises: the user-side constraint at the Fable capability level.

Cross-links

Source