# 将大语言模型的输出"人性化"是愚蠢的

- 来源：Hacker News 热门（buzzing.cc 中文翻译）
- 作者：kuberwastaken
- 发布时间：2026-08-11 06:02
- AIHOT 分数：62
- AIHOT 链接：https://aihot.virxact.com/items/cmsnshmyu0iborohfxy7msqbu
- 原文链接：https://kuber.studio/blog/Reflections/Humanising-LLM-Outputs-is-Actually-Dumb

## AI 摘要

给 LLM 输出套上“人性化”风格指令（如“我有 ADHD”或 ASD-STE 简化技术英语）是错误抽象：这类压缩是有损的，会掩盖失败细节，且让智能体间传递信息时层层丢失关键状态。作者主张智能体应默认保留高保真机器状态（如精确错误、堆栈、置信度），仅在人类消费边界做压缩渲染。

## 正文

Humanising LLM Outputs is Dumb

The largest tell for me to tell where culture and sentiment is shifting for AI tools is usually X, viral GitHub repositories and Hacker News.

One of these tells I’ve been seeing a lot lately is skills like I have ADHD and Agents.md instructions such as giving outputs in only ASD-STE100 Simplified Technical English.

I understand the appeal, none of us really like the verboseness and specific quirks of LLM outputs, but I really think fixing that by humanising the model is the wrong abstraction.

The problem is that these instructions are not applied after the model has finished doing the work, it becomes part of the same work - If you tell an agent to use short sentences, avoid jargon, never overwhelm you and only include the most important details, you are asking it to continuously compress its output into a lower-bandwidth format.

That compression is lossy.

You probably never notice what got dropped because the output still reads nicely.

ASD-STE is a great example because it sounds so reasonable. It was designed to make documentation unambiguous for humans. But an agent isn’t a human technical writer, and the raw state is often the most information-dense representation available. Meanwhile the style rules sit on the same instruction list as: solve the task, use tools correctly, preserve abstractions, don’t break anything.

This becomes even stranger once agents start talking to other agents.

A subagent investigates a bug, turns its findings into a nice human-readable summary, the parent agent reads that summary, and then turns it into another nice human-readable summary for you.

If a subagent ran six tests, I don’t want:

Most tests passed, although there was one issue worth looking into.

I want:

5/6 PASS FAIL: test_cache_invalidation CAUSE: stale key survives restart REPRO: tests/cache_test.py:184

More importantly, humanisation hides failure.

Agents fail in useful, ugly ways: conflicting evidence, unresolved branches, stack traces, uncertain assumptions. Human prose is extremely good at smoothing these into sentences like:

There are a few considerations here.

That sounds nicer.

But I’d rather find my agent is hallucinating or near its token window than be happy with that.

Every other system we build works the opposite way - Databases don’t store data in the format a dashboard displays it, compilers don’t make their IR pleasant to read, APIs don’t exchange friendly summaries.

We keep the highest-fidelity representation as long as possible and transform it at the boundary where a human consumes it, but LLM tooling is increasingly doing this backwards.

To be clear, none of this is an argument against accessibility or personalisation.

If you want three-line answers or Simplified Technical English, great! I just think it’s better to do it at the end.

Let agents keep detailed state, let subagents exchange schemas, diffs, exact errors, confidence, provenance. Then compress it for me.

I think the best part is that these viral skills might actually be pointing toward the right future.

Users are patching this at the prompt layer, something that belongs further down the stack.

“Talk to me like I have ADHD” makes perfect sense as a renderer, it makes much less sense as an operating instruction. The durable version is agents whose native language is precise, machine-facing state, with the warm, concise, human version generated only at the boundary.

So the viral repos aren’t the end state, but a bug report.