Thariq Shihipar

Sources#

• A Field Guide to Fable: Finding Your Unknowns
• Fable's judgement
• How I AI: Thariq Shihipar on Replacing Markdown with HTML for AI-Powered Development
• The new rules of context engineering for Claude 5 models

Summary#

Engineer on the Claude Code team at Anthropic. Source of the "HTML is the new markdown" thesis (see HTML as the New Markdown) and the Compute Allocator framing of the human role. Interviewed by Claire Vo for the "How I AI" series after the two met at Code with Claude in San Francisco (episode published May 2026). Personal site thariq.io; on X as @trq212.

Notable claims and stances#

• "HTML is the new markdown." As models get more capable and projects more complex, markdown plans overwhelm the human reader; HTML is a richer medium that keeps humans engaged. The model isn't the bottleneck — human attention is. See HTML as the New Markdown.
• We're all becoming "compute allocators." The human's job is deciding what's worth spending compute on; ~1% of his generated tokens reach production, the other 99% is scaffolding (plans, interfaces, status, design systems) invested in alignment and communication. See Compute Allocator.
• Abundance mindset. Generation is cheap enough to afford throwaway tooling — micro-apps built for a single editing task ("micro-software on top of micro-software").
• Constraint plus latitude prompting. "Provide enough constraint to get what you want, but leave room for the model to surprise you" — "Hey, Claude, I trust you here."
• Recenter the human. His workflows are explicitly not about automating the developer away; they make collaboration more visible and engaging.
• The map is not the territory. (July 2026 field guide.) Your prompts, skills, and context are the map; the codebase and its real constraints are the territory; the gap is unknowns. "Fable is the first model where I find the quality of the work is bottlenecked by my ability to clarify its unknowns." Reducing and planning for unknowns "is the skill of agentic coding" — and it is learnable. See Unknowns as the Agentic Bottleneck.
• The quiz gate. "I only merge after I pass the quiz perfectly" — he has Claude quiz him on a change before merging, because reading diffs under-informs when behavior depends on existing code paths. The human, not the agent, is what gets verified.
• Source code is the best reference. When you can't describe what you want, point at code that does it — even in another language. This is how Claude Design works: point it at a component and "it reads the underlying code, not just the screenshot."
• Unhobbling. (July 2026 context-engineering post.) The Claude Code team was over-constraining the model: they deleted >80% of the system prompt for Claude 5 models with no measurable eval loss, and retired a set of context-engineering "myths" (rules→judgment, examples→interface design, upfront→progressive disclosure, repetition→tool descriptions, CLAUDE.md-memory→auto-memory, simple specs→rich references). Shipped to users as claude doctor. See Harness Shrinkage as Models Improve and Agent Context Files. He and Cat Wu gave the rules→judgment half as user-facing advice at an AI Engineer World's Fair fireside — let Fable "use their own judgement rather than dictating how they should work," with automated testing as the worked example (2026-07-03, relayed second-hand in Fable's judgement).

The three workflows he demonstrated#

1. Interactive HTML brainstorming and planning — ask for plans as HTML files, not lists; visual, scrollable, shareable.
2. Disposable Micro-Apps — generate a throwaway custom UI to edit one part of a plan, with copy-back-to-markdown.
3. Living Design System — a design_system.html extracted from repos as a portable, human- and machine-readable source of truth.

Connections#

• Claude Code — his team at Anthropic
• Anthropic — employer
• Claire Vo — interviewed him for "How I AI"
• HTML as the New Markdown — his central thesis
• Compute Allocator — his framing of the human role
• Disposable Micro-Apps / Living Design System — his other two workflows
• Boris Cherny — Claude Code teammate; "coding is solved (for me)" is the precondition for Thariq's allocator framing
• Harness Shrinkage as Models Improve — Thariq's 99%-scaffolding observation is the human-facing counterweight to the team's model-facing harness pruning
• Engineer PM Convergence — his compute-allocator framing is "deciding what to write" exercised at invocation scale
• AI Native Product Cadence — his richer-HTML-PRD bet is the counterpoint to Cat Wu's lighter-PRD bet
• Interaction Models — both target better human–AI collaboration; his HTML artifacts enrich the asynchronous channel where TML dissolves the real-time interface
• Software 3.0 — his HTML-first workflows and Disposable Micro-Apps are Software-3.0-native: per-task UIs spun up from a prompt rather than hand-built apps
• Unknowns as the Agentic Bottleneck — his July 2026 thesis: with Fable-class models the binding constraint is the human's ability to articulate what they never wrote down
• Context Advantage, Not Taste — his unknown knowns are Andrew Ng's context advantage seen from inside the human's head; the field guide is the elicitation protocol for it
• Claude Fable 5 — the model that, in his account, moved the bottleneck onto the human

Sources#

• How I AI: Thariq Shihipar on Replacing Markdown with HTML for AI-Powered Development — "How I AI" interview with Claire Vo (ChatPRD), May 2026
• A Field Guide to Fable: Finding Your Unknowns — "A Field Guide to Fable: Finding Your Unknowns," published on X, 2026-07-04 (practitioner-opinion)
• Fable's judgement — Simon Willison, 2026-07-03 (practitioner-opinion): the AI Engineer World's Fair fireside advice, relayed second-hand and attributed jointly with Cat Wu without saying who said what; no transcript or recording is in the corpus