Humanizing AI output is the wrong layer — render at the boundary, not mid-task

A wave of viral prompt tricks — agent instructions like “talk to me like I have ADHD” or “respond only in Simplified Technical English” — tries to tame the verbosity and quirks of LLM output. The author’s objection isn’t to the goal but to where it’s applied. These style rules don’t run as a post-processing step; they sit on the same instruction list as “solve the task, use tools correctly, don’t break anything.” Telling an agent to stay short and jargon-free forces it to continuously compress while it works, and that compression is lossy. Because the result still reads well, you never see what got dropped.

The problem compounds in multi-agent setups. A subagent that ran six tests should hand the parent exact results, errors, and confidence levels — not a tidy sentence like “most tests passed, with one issue worth looking into.” Human-friendly prose is especially good at smoothing over the ugly signals that actually matter: conflicting evidence, unresolved branches, stack traces, shaky assumptions, or a model drifting toward hallucination near its context limit. Humanization, in other words, hides failure.

The author frames this as backwards from every other system we build: databases, compilers, and APIs all preserve the highest-fidelity representation internally and transform it only at the human-facing boundary. The fix is to let agents keep detailed machine-facing state and exchange schemas, diffs, exact errors, and provenance — then compress to a warm, concise form only at the point a person reads it. Personalization and accessibility are fine; they just belong at the render step. The viral “humanize me” skills aren’t the destination but a bug report, signaling that users are patching at the prompt layer something that really belongs lower in the stack.