The 35-word rule: prompting an LLM whose output will be spoken aloud

The first version of this site's voice clone was a great writer and a terrible conversationalist.

Asked "what did you build at InCruiter?", it produced a lovely, accurate, well-structured answer of 69 to 88 words. On a screen that's a short paragraph. Spoken by a TTS, it's about 38 seconds of uninterrupted talking. Nobody on a call waits 38 seconds. They interrupt, or they leave.

Text LLMs are trained to be thorough. Voice needs the opposite. Here is what changed.

Rule 1: a hard word limit, stated as a hard limit

"Be concise" does nothing. Every model believes it is already concise. What works is a number:

HARD LIMIT: one or two short sentences, max 35 words, one paragraph.
Give the single best point, then offer more.

The second line matters as much as the first. A voice answer isn't a shorter essay, it's the first move in a conversation. Give the best point, then ask "want me to go deeper?" The caller steers. That's what a person would do.

Rule 2: don't trust the prompt alone

Models follow a word limit most of the time. "Most of the time" is not a product. So the agent also enforces it server-side, inside the text-to-speech step (code):

async def replay():
    spoken = head
    yield head
    async for chunk in chunks:
        yield chunk
        spoken += chunk
        if len(spoken.split()) > MAX_SPOKEN_WORDS and spoken.rstrip()[-1:] in ".?!।;":
            return  # past 45 words: stop at the next sentence end

Two details:

• It cuts at a sentence boundary, never mid-word. A clipped sentence sounds like a dropped call.
• The cap (45) sits above the prompt's limit (35). The guard is a backstop for when the model rambles, not the main control.

Note the ।: that's the Devanagari full stop. A sentence splitter that only knows .?! will happily run on through an entire Hindi paragraph.

Rule 3: strip everything that isn't speech

LLMs love markdown. TTS engines read it literally. Without a filter, "important" can come out as "asterisk asterisk important asterisk asterisk".

The prompt says no markdown, lists, emojis, URLs or code, and the session strips them anyway before the text reaches the voice:

tts_text_transforms=["filter_emoji", "filter_markdown"]

Belt and braces. You will not catch every way a model can format text, so remove formatting at the last step too.

Rule 4: numbers the way people say them, but exact

The model once turned "1,000+ concurrent interviews" into "thousands of interviews". Sounds more natural. Also no longer true. On a portfolio, that's the difference between a fact and an exaggeration.

The rule now: say numbers the way people say them, but quote them exactly. Natural delivery is the TTS's job. Accuracy is the prompt's.

Rule 5: the model will invent your life story

This one was embarrassing. A test caller asked how I got my job. The model, trying to be helpful and personal, replied that I got an internship through my open-source pull requests.

Lovely story. Never happened.

Voice makes this worse than text. A spoken sentence in a confident first-person voice sounds like testimony. Listeners don't scroll back and fact-check. So the prompt now has an explicit rule plus examples:

Make up personal stories or outcomes. Don't say how you got a job or
internship, what PRs led to, or anything you "did" unless KNOWLEDGE says it.
Give advice as advice ("what works is..."), not as invented history.

The examples teach the difference between "how did you do X?" (answer: that's not on my portfolio, here's a tip) and "how do I do X?" (answer: advice). Temperature went down to 0.4 too: lively enough to sound human, low enough to stop embellishing.

Rule 6: examples beat instructions, but examples are contagious

Examples fixed the invented stories and then caused a new bug: an English question got a Hindi answer, because the examples were mostly Hindi and the model copied the language along with the behaviour.

The fix was to make the examples cover each language and each behaviour, so the model learns the rule rather than the surface. Examples are the strongest signal in a prompt. They teach everything that's in them, including what you didn't mean to teach.

The checklist

The whole prompt is on GitHub, and the agent it drives is on the home page. Try asking it something it shouldn't know.