Thariq Shihipar

Member of technical staff at Anthropic, works on Claude Code. Ex-founder of a YC startup (~30 people). X: @trq212.

Contributions in this vault

Field Guide to Fable (Thariq Shihipar, Anthropic) (2026-07-03 WF talk) — introduces the vault's canonical formulation of Capability Overhang (via the Pokémon-AW example), the unknowns matrix for finding what you haven't specified, and the "only way out is through" frame for the practitioner grief of Fable-class model transitions.
The New Rules of Context Engineering for Claude 5 (Thariq Shihipar) (2026-07-24 Anthropic blog) — official-artefact version of the same thesis: six named pattern shifts (rules→judgment, examples→interfaces, upfront→progressive, repetition→simple, manual-memory→auto, simple-specs→rich-references) plus Anthropic's claude doctor / /doctor tool for rightsizing existing skills and CLAUDE.md files. See Context Rightsizing for the discipline this codifies.
The ask-user-a-question tool. He built it. Its progression — barely working in Opus 4, interviewing users in Opus 4.5, generating embedded-HTML questionnaires in Opus 4.8 / Fable — is one of the clearest concrete traces of the capability overhang.
First-party evidence for the 80% Claude Code system-prompt reduction for the new model class — models more imaginative than examples, so examples now constrain. First delivered on stage in July, formalised as published Anthropic guidance on July 24.

Anthropic-cultural voice. Delivers the "tradeoffs are not real" frame from inside the company; delivers the "remove don't add" frame in the blog.
Convergent with Matt Pocock — his same-conference talk arrived at the same interview-the-human-before-code prescription via Design Concept (Brooks). Both cite Brooks explicitly. Two-source convergence on the same practical move, one week apart. The July 24 blog extends the convergence — Anthropic's "remove don't add" + /doctor is the first-party version of Pocock's deletion test in Skill Checklist (Pocock).
Practitioner voice on Mythos-class transitions in the vault — pairs with Nate B. Jones (task imagination, TikTok-scale), Nathaniel Whittemore (advantage gap, podcast-scale), and Alex Albert (Anthropic-launch positioning).
Second substantive artefact in three weeks (WF stage → official blog) makes Shihipar the vault's densest first-party voice on how Anthropic itself now recommends structuring prompts / skills / CLAUDE.md for the current model class.