AI authorship timing capture

Claude Opus 4.8 round trip vs the generated Heat path.

One prompt asked Claude Code CLI to write a small Heat program. The model round trip was live and variable; the generated Heat compile/run path was local and deterministic.

Live LLM authoring

Claude Opus 4.8

via Claude Code CLI

76.94s

round trip, variable, not a guaranteed SLA

Local Heat path

Generated Heat program

compiled locally

0.05s

compile, plus <0.01s run

Manual edits0 to the generated Heat program.

Runtime<0.01s for the compiled binary.

Claim boundaryLLM timing is an observation, not a benchmark or SLA.

Why This Matters

The goal is not to make the model faster.

The goal is to need fewer model round trips. Heat is designed so AI-authored code has fewer ambiguous choices and more compile-time refusal when the first draft is wrong.

Artifacts

The raw response included a short prose preface before the Heat code, so this page does not claim the full terminal transcript was a valid Heat source. The generated Heat code itself was compiled unchanged.

prompt.txt claude_response_raw.txt claude_response.heat capture.txt results.json social_card.svg social_card.png

Measured locally via Claude Code CLI and /usr/bin/time -p. LLM timing includes client, network, provider, and model latency.