How do you make an open model sound human without making it dumber?
Open-weight models answer correctly but sound stiff and list-heavy. Telling them to talk casually makes the tone better, but the reasoning gets sloppier and they start skipping steps. Is there a way to get a natural voice and keep the quality of the thinking?
Split the job in two. Asking one prompt to reason carefully and also sound relaxed puts two goals in competition, and smaller models usually trade away the reasoning first. Let the model work out the answer in its normal style, then run a second pass whose only job is to rewrite the finished answer in a human voice.
The rewrite pass works best with concrete rules instead of adjectives. "Sound natural" is vague; "no bullet lists unless asked, no opening praise, no 'it's worth noting', vary sentence length, keep every fact and number" is something the model can follow and you can check.
Give one or two short examples of the voice you want. A sample paragraph teaches tone faster than a page of instructions, and with open models it is often the single biggest improvement.
Check the rewrite against the original before trusting it. Compare the facts, numbers and steps in both versions on a handful of real answers; if the rewrite drops or changes anything, tighten the rule that says meaning must not change.
Listings mentioned
- Humanizer · skill by manojbajaj95A rewrite pass that removes common AI writing tells while preserving meaning and voice, the second step of the split described above.
- Prompt engineering patterns · skill by wshobsonCovers few-shot examples and structured prompting, the tools for teaching a voice with samples instead of adjectives.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.