Q&A: Thariq Shihipar and Simon Willison on HTML as AI Output Format¶

Thariq article: thariq-shihipar-unreasonable-effectiveness-html
Simon article: simon-willison-unreasonable-effectiveness-html

Q1: What is Thariq arguing that Simon agrees with?¶

Both agree on the core thesis: Markdown is the wrong default for AI coding agent output.

Thariq's specific claims Simon endorses:
- HTML artifacts are richer — color-coded severity, collapsible sections, inline diagrams, margin annotations all beat text labels
- HTML feels like a starting point; Markdown feels like an endpoint — one invites continuation, the other invites filing
- The handoff problem is real — the hardest part of agentic work isn't execution, it's transferring context to a human who needs to act on the result. HTML bridges that better than Markdown
- Token efficiency no longer justifies Markdown — the 8,192 context window constraint that made Markdown dominant is gone

Simon took Thariq's argument and ran with it: he tried the curl | llm -s 'explain in HTML' pattern live on a Linux exploit and produced a result he describes as "pretty good." The experiment convinced him to shift his default.

Q2: How does Simon extend or refine Thariq's argument?¶

Simon adds three things Thariq doesn't emphasize:

1. The mechanism matters: self-contained, no hosting
Thariq focuses on the capability. Simon adds the practical delivery mechanism: single-file HTML with inline CSS and JS, opened locally in a browser. No server, no build step, no deployment. The artifact just works when you open it.

2. The copy.fail pattern as a repeatable workflow
Simon demonstrates a specific CLI pattern — curl URL | llm -s 'explain in detail... output HTML' — that any developer can adopt immediately. This makes HTML output a protocol, not just a capability.

3. The agentic engineering connection
Simon frames HTML artifacts as the natural extension of the "write code is cheap now" insight. Rendering rich output is also cheap now. The bottleneck isn't generation — it's human comprehension speed. HTML solves comprehension the way code generation solved execution.

Q3: What does this exchange tell us about how AI researchers are thinking about LLM output formats?¶

Three patterns stand out:

a) The artifact is the interface
Thariq's framing is explicit: "Markdown is a report. HTML is an interface. Reports are for reading. Interfaces are for continuing the work." This is a significant conceptual shift — from treating AI output as a document to treating it as a tool.

b) Internal practice before external argument
Thariq works on Claude Code at Anthropic. The fact that his team is producing HTML artifacts as part of their internal workflow suggests this isn't speculative — it's operational. Simon validated Thariq's argument by running it himself rather than treating it as theory.

c) The format debate is really a cognition debate
Both Thariq and Simon are ultimately arguing about how to reduce cognitive load on the human receiving the output. The format is a proxy for comprehension speed. This maps directly to factory QA — the question isn't "what format?" but "what gives the operator the fastest path to a decision?"

See also: simon-w-agentic-engineering-patterns for the broader agentic engineering context, and kelly-factory-overview for factory quality gates where HTML artifacts could replace Markdown diffs.