Kuber Mehta · 2026-08-10 · notable
Kuber Mehta — 'Humanising LLM Outputs Is Dumb'
Kuber Mehta argues that the popular skills telling an agent to write more like a human are a design mistake. Making a model reformat its findings into friendly prose is lossy compression that throws away the failure signals you needed.
Telling an agent to sound human makes it compress away the exact failure details you needed to debug.
What is it?
"Humanising LLM Outputs Is Dumb" is an August 10, 2026 blog post by Kuber Mehta on the popular agent skills that instruct a model to write in a plainer, more human register — the "I have ADHD" prompts and the ASD-STE100 Simplified Technical English rules doing the rounds. Mehta's claim is that these fix a symptom in the wrong place.
How does it work?
The mechanism Mehta describes is lossy compression. Asking a model to finish a task and reformat its report at the same time makes it squeeze real findings into readable prose, and detail falls out. His example contrasts an agent saying "Most tests passed, although there was one issue worth looking into" with the machine-readable version — "5/6 PASS / FAIL: test_cache_invalidation / CAUSE: stale key survives restart / REPRO: tests/cache_test.py:184" — which names the test, the cause, and the line to reproduce it.
Why does it matter?
The fix Mehta proposes is architectural: keep the highest-fidelity representation for as long as possible and transform it only at the point a human reads it, the way databases, compilers, and APIs already work. For anyone wiring agents into CI or a review pipeline, that reframes the viral humanising skills as useful bug reports about output format rather than the answer. The post reached the Hacker News front page with 185 points and 114 comments.
Who is it for?
developers building agent pipelines and tooling