Yoav Goldberg floats 'harness distillation' to explain Claude Code's 80% shorter prompt
yoavgo · x · 2026-09-04
Researcher yoavgo (Yoav Goldberg) proposes a "harness distillation" hypothesis in reply to "maybe the models just became good enough":
- Boris reportedly said Claude Code's prompt is now 80% shorter — plausibly because models were trained to follow the instructions of the longer prompt, letting the external prompt shrink.
- He extends this to Claude harness "reminders" being removed: models could be trained to follow plans without reminders using outputs of models that did get reminders.
His point: harness simplification may reflect distilling harness instructions into training rather than raw capability gains.
More from coding & agent
- Cohere Labs to host session on building voice agents for Asian languages — Cohere_Labs · 2026-09-04
- JetBrains deep dive: Go error handling techniques, best practices and common pitfalls — rseroter · 2026-09-04
- Agent training series finale: building agent environments with trl and openenv — SergioPaniego · 2026-09-04
- Open-source benchmark targets the tool layer: auth, retries, memory failures that agent evals miss — Kind-Atmosphere9655 · 2026-09-04
- Rust values vs places: the mental model that makes ownership click — Abhishekcur · 2026-09-04
- Claude Code on iOS is getting slash command support — testingcatalog · 2026-09-04