Thariq Shihipar's AI Engineer talk: unhobbling Claude, capability overhang, finding unknowns, system-prompt shrink, copy-paste Fable prompts, grief over hand-coding, and being unreasonable with frontier models.
At AI Engineer, Thariq Shihipar — member of technical staff at Anthropic on Claude Code — opened with a Claude Code tradition: a pre-talk selfie with the room. Then the headline: Fable is back, rolling out later that day, with a 12:30 fireside alongside Kat Wu and Simon Wilson for possible updates.
The core of the session was a four-part field guide Thariq had been publishing as articles — compressed into one stage speed-run:
Anthropic removed ~80% of Claude Code system prompt for Fable-class models
Core risk with Fable?
Map ≠ territory — your spec is not the codebase; unknowns compound at scale
Top tactic?
Blind spot pass before implementation
Emotional angle?
Grief for hand-coding; only way out is through
Ambition angle?
Be unreasonable — force tradeoffs to reveal themselves
Part 1 — Unhobbling Claude
Thariq's opening frame: models are grown, not designed. Anthropic does not wake up targeting 99% on SWE-bench — teams grow models with data, feedback, and compute, then learn alongside them. What contains a model in production is us: the harness, tools, and prompts.
Unhobbling means understanding Claude better so you stop leaving capability on the table.
Capability overhang (the Pokémon test)
A viral puzzle: Which Pokémon names end in "AW"? (~1,000 species; answer: two — Croconaw and Drednaw).
Surface
Behavior
Chat-only model
Often fails — knows names but cannot reliably enumerate
Claude Code
Fetches data, writes a filter script, returns both names
The model did not suddenly memorize every species. Code execution unlocked a spiky skill. Thariq calls this capability overhang — intelligence present but invisible until the harness exposes it. Fable's challenge is discovering what overhang exists now.
How harnesses evolved (three spikes)
1. Context → arms (birth of Claude Code)
Early coding intuition: paste the whole repo into a giant context window. What worked instead: give the model bash, search, and environment access so it builds its own context. That insight is Claude Code itself — not bigger prompts, better arms.
Claude Code waits for you to prompt. Claude Tag pushes toward agents that wake themselves up and work in multiplayer settings — another harness shift, not just a smarter weights upgrade.
If your CLAUDE.md and skills still compensate for a weaker model, you may be hobbling Fable with obsolete scaffolding.
Ask-user-questions: another overhang curve
Thariq built the ask user a question tool early at Claude Code:
Model generation
Behavior
Opus 4
Barely called the tool — required heavy tuning
Opus 4.5
Could interview you across dozens of spec questions
Opus 4.8 / Fable
Builds HTML reports with embedded question flows
Same tool name; radically different UX as capability jumps. See HTML vs Markdown in Claude Code for the output-format parallel.
Biology, not physics
Thariq urges treating Fable like biology — empirical, organic, rules incomplete — not a physics equation. He pointed attendees to Anthropic's research essay Biology of a Large Language Model for intuition-building.
Part 2 — Finding your unknowns (unhobble yourself)
Models improved; you must improve too. Thariq's map/territory frame:
Map — your plan, prompt, spec, mental model
Territory — real codebase, constraints, production reality
Unknown — territory not on your map; Fable must improvise
Fable was the first model where Thariq felt he had to catalog unknowns — autonomous reach is so wide it hits unstated decisions fast. Deeper analysis: map is not the territory guide.
The four-quadrant matrix
Known
Unknown
Known
Known knowns — what you put in the prompt
Known unknowns — gaps you know you have not researched
Unknown
Unknown knowns — "know it when you see it" (design taste)
Unknown unknowns — blind spots you have not considered
Fable can help surface all four — if you prompt for discovery before execution.
Copy-paste prompts from the talk
Blind spot pass (unknown unknowns):
text
I'm working on a new auth provider I know nothing about in this codebase.
Can you do a blind spot pass to help me figure out my relevant unknown
unknowns and help me prompt better?
Point Fable at git diff, Slack, or docs for richer maps. Thariq used the same pattern learning color grading for video — domain-agnostic.
Design prototypes (unknown knowns):
text
I have no visual taste. Make me an HTML page with four widely different
design decisions so I can react to them.
Reaction beats description when you cannot verbalize taste.
Architecture interview (known unknowns):
text
Interview me about this feature. Prioritize questions that would change
the architecture. [Add context about stage, constraints, and audience.]
Reference as map (skip writing the spec from scratch):
text
Here's code that represents what I want done — possibly another language
or system. Read it, understand it, and use that as your starting map.
Works with HTML mockups as references for React components — ties to Thariq's HTML thesis.
Implementation notes (mid-run unknowns):
When Fable hits an unspecified decision, ask it to log the deviation so you can see where the map failed and update CLAUDE.md or skills permanently.
Quiz before merge (stay in the loop):
Have Fable quiz you on what changed before you open a PR — ensures you can represent the work, not just approve diffs.
Staying in the loop
Thariq's recurring theme: Fable's output scale demands human alignment. Autonomy without map maintenance produces confident wrong paths. Pair these tactics with loop engineering so sessions end with updated territory notes, not one-off fixes.
Part 3 — Dealing with the grief
The talk pivoted personal. First Mythos-class session brought gain and loss.
Thariq ran a ~30-person YC startup where code forced brutal tradeoffs — ship fast or prototype features; weeks vs months per bet. Revisiting that codebase with Fable: work that took weeks became hours.
He loves hand-written code — the mental rotation of a system in mind. He also remembers late-night debugging, weeks without progress, most projects failing, most startups dying. Programming was hard; the highs were real; he cannot go back.
His resolution: the only way out is through. Grieve the craft; keep learning agentic coding with Fable while staying in the loop. Sentiment echoed in Fable loop design — longer autonomy changes your job, not just the model's.
Part 4 — Being unreasonable
Anthropic culture, per Thariq: tradeoffs are often fake. At his startup he "was reasonable" — quarterly priority lists, explicit sacrifices. At Anthropic the question is: what if you tried to do all of it and let reality expose the real constraint?
Good, fast, cheap — pick three softens when Fable compresses execution cost. Thariq built his AI Engineer deck in ~4 hours with Fable — quality he liked, speed he did not expect.
His challenge to the room: the world is watching AI engineers prove this is not a fad — more productive, work less, spend time with people you care about.
Caveat he named:Building is easier; generating value is still hard. Optimizing harnesses and tooling is not the goal — swings at valuable outcomes are. Connects to context vs prompt vs loop vs harness: the stack serves outcomes, not the reverse.
Fable availability context (July 2026)
The talk assumed Fable rolling out imminently — consistent with Anthropic's July 1 global restore after the June export-control suspension.
Subscription note: included Fable on Pro/Max/Team ends July 7, 2026, 11:59:59pm PT — then usage credits. Anthropic still pledges eventual return to subscription inclusion when capacity allows. Details: subscription return guide.
Harness advice in this recap applies regardless of billing bucket — but credit economics may push you toward Sonnet 5 for routine map-building and Fable for high-unknown passes.
What to try this week
Audit your harness — remove skill scaffolding written for pre-Fable models; test with a smaller effective system prompt.
Run one blind spot pass on the hairiest module you are about to touch.
Log implementation notes for one long Fable session; convert surprises into CLAUDE.md updates.
Pick one "unreasonable" project — scope you would have cut last year; time-box with Fable and measure value, not just LOC.
Recap based on Thariq Shihipar's AI Engineer stage talk and his published field-guide articles. Fable availability, system prompts, and Claude Code defaults change frequently — verify in your CLI and Anthropic docs before treating any prompt as permanent.
