Thariq Shihipar

Thariq Shihipar, a member of the équipe Claude Code at Anthropic, appears in four source notes across two months, all of them about the same problem: what a human still has to supply once the model is good enough.

His May 2026 argument is a format switch. Markdown, he writes, becomes restrictive past roughly 100 lines, and since he no longer hand-edits his specs (he prompts Claude to edit them), Markdown's editability advantage falls away. HTML replaces it as the default agent output: tables, CSS, SVG, sliders, single-file throwaway editors that end with a "copy as prompt" button. He does not hide the cost. HTML diffs are noisy and hard to review, and he leaves that unresolved. He also refuses to package the practice as a skill, saying it is too contextual to freeze.

In June, writing on Anthropic's blog in «Lessons from building Claude Code: How we use skills», he inverts the usual documentation instinct: do not write down what Claude already knows. The Gotchas section, built from real failure points, is the highest-signal content in a skill.

By July his framing has sharpened into a claim about the bottleneck: "Fable est le premier modèle où la qualité du travail est plafonnée par ma capacité à clarifier ses inconnues." Reducing and planning for your unknowns, he argues, is THE skill of agentic coding. The constraint has moved from the model to the person describing the work.