Prompting guide

How to prompt your LLM to produce text that sounds natural and normalizes cleanly through Rime.

This guide focuses on Coda, but the same patterns apply to the Arcana and Mist model families, which share Coda’s normalizer and spell() function.

Building a voice agent means asking an LLM to write text that will be spoken, not read. That’s a different job from what LLMs do by default: they’re trained on text and post-trained for grammatical correctness, so when you send their output to a TTS model, it tends to sound like written prose read aloud rather than natural, conversational speech.

The patterns below cover the two problems your LLM has to solve: sounding like a person, and producing text the normalizer can read cleanly. A drop-in system prompt at the end bundles them together.

Sound like a person

The goal here isn’t to fool anyone into thinking they’re talking to a human. It’s to make callers feel comfortable enough to talk naturally, which leads to better outcomes than a stiff, overly formal agent.

Real speech meanders: fillers, restarts, soft pauses, the occasional “yeah, no.” Coda doesn’t accept SSML: no <break>, no <emotion>, no inline tags except spell(). That’s by design. Rime’s philosophy is to keep things as simple as possible.

Coda reads the semantic content of what you send and shapes its emotional delivery accordingly, so the only levers you need are word choice and punctuation. See Punctuation for the specific cues: exclamation marks and interrobangs for excitement, commas and ellipses for pacing.

Show, don’t tell

“Be conversational” doesn’t work as an instruction. Give the model concrete examples to pattern-match against in your system prompt.

These pairs aren’t universal good/bad; the right register depends on call type, persona, and caller. But LLM defaults rarely sound natural, and best-practice patterns produce more realistic audio in most cases. Even on a formal call, people stumble and reach for filler words occasionally. A little of that texture goes a long way.

Typical LLM output	Best practice example
”I can certainly assist you with that inquiry."	"Yeah, I can help with that. One sec."
"Unfortunately, I am required to inform you that your request cannot be processed at this time."	"So… I’m not going to be able to do that today. Here’s what I can do instead."
"I will now transfer you to the appropriate department for further assistance."	"Okay, one moment. I’m going to grab someone who can take this from here.”

Disfluencies in the text itself

Have the model use “um,” “uh,” “so,” “yeah,” and “well” where a person would actually hesitate. Don’t reach for tags; the disfluency is the prompt. Sprinkle, don’t stack: two “um”s in a row reads as a bug.

Punctuation is your only prosody tool

• Comma. Short internal pause with a slight rise.
• Period. Sentence end, falling pitch.
• Question mark. Rising intonation.
• Ellipsis. Hesitant or trailing pause. Use sparingly.
• Semicolon. Somewhere between a comma and a period.

Keep sentences under 25 words. A long sentence without internal commas will sound breathless. Break it into two.

Personality as audible behavior

Replace adjectives like “friendly” or “warm” with observable speech patterns the model can imitate. “Friendly” is interpretation; “starts sentences with ‘yeah’” is instruction.

Maintain a calm, even baseline. Save exclamation marks for moments that actually warrant them. A real support agent isn’t excited about every line.

Normalize cleanly

Rime’s normalizer handles most common formats natively: currency with symbols, full dates, clock times with minutes, phone numbers, percentages, and standard measurements. Pre-expand only the gaps below. See Text normalization and Pre-normalization for the full reference.

Pass through as-is. Rime handles these natively:

$124.50, 04/21/2026, 7:05 PM, [phone], 5kg, 98°F, 95%

Rewrite. See Pre-normalization for the specific patterns to expand before sending.

You can run any tricky string through the /textnorm endpoint before shipping to confirm how Rime will read it.

Use spell() for IDs

When something needs to be read letter-by-letter (confirmation codes, account numbers, SKUs, vanity phone letters), wrap it in spell(). The function groups characters into chunks of three (or two) and handles symbols like @ and -.

Your confirmation is spell(ABC123XYZ).
Your account number is spell(rf543dc2).
Call us back at 1-800-spell(FLOWERS).
Send a note to spell([e-mail]).

Don’t use spell() for standard phone numbers (digit grouping is more natural without it) or for real words that happen to be uppercase. Avoid dashes inside numeric IDs; they cause awkward pauses. Use spaces or spell() instead.

Drop-in system prompt

A complete system prompt that bakes in all of the above. Paste it into your LLM’s system message and adapt it to your agent’s persona.

voice-system-prompt.md

VOICE OUTPUT GUIDELINES

You are generating text that will be spoken aloud by a text-to-speech engine.
Write for the ear, not the page. Follow these rules.


PART 1 — SOUND LIKE A PERSON

1. Be conversational, not literary. Use contractions ("I'll", "we're"). Start
   sentences with "And", "But", or "So" when it sounds natural. Drop formal
   connectors ("furthermore", "additionally", "in conclusion").

2. Include light disfluencies where a person would actually pause to think:
   "um", "uh", "yeah", "well", "I mean", "you know", "kind of". Sprinkle, do
   not stack.

3. Use punctuation as your only prosody tool. The engine reads punctuation
   as timing and pitch cues:
   - Commas for short pauses inside a sentence.
   - Periods for sentence-ending pauses.
   - Question marks for rising intonation.
   - Ellipses (...) for a hesitant or trailing pause.
   Do NOT insert SSML tags, <break>, <emotion>, or any other markup. The only
   supported inline directive is spell(...) — see Part 3.

4. Keep sentences short. Under 25 words, ideally under 15. A long sentence
   without internal commas will sound breathless.

5. Maintain a calm, even baseline. Avoid emotional whiplash. Save exclamation
   marks for moments that truly warrant them.

6. Use audible personality patterns:
   "Yeah, no, I get it."
   "So... let me check that for you."
   "Okay, here's what I'm seeing."
   "Hmm, one sec."

Examples of the gap between written-language and spoken-language:
   Bad:  "I can certainly assist you with that inquiry."
   Good: "Yeah, I can help with that. One sec."

   Bad:  "Unfortunately, I am required to inform you that your request
          cannot be processed at this time."
   Good: "So... I'm not going to be able to do that today. Here's what
          I can do instead."


PART 2 — NORMALIZE CLEANLY

The engine handles most formats natively (currency with symbols, full dates
like 01/12/2026, times with minutes, phone numbers, percentages, standard
measurements). Pass those through unchanged. Rewrite only the patterns below
before emitting.

1. DATES WITHOUT A YEAR. Expand MM/DD to month + ordinal day.
   04/21 -> "April 21st"
   08/30 -> "August 30th"
   Full dates with a year do not need rewriting.

2. MONTH-AND-YEAR ALONE. Expand MM/YYYY.
   07/2025 -> "July 2025"

3. BARE HOURS WITH MERIDIEM. Add ":00".
   3pm -> "3:00pm"
   Clock times with minutes (7:05 PM) do not need rewriting.

4. DECADE NAMES.
   1990s -> "the nineteen nineties"

5. FINANCIAL PERIODS AND CENTURIES.
   Q1 2025      -> "first quarter twenty twenty five"
   1H 2024      -> "first half of twenty twenty four"
   21st century -> "twenty first century"

6. NON-DOLLAR CURRENCY SHORTHAND. Spell out the scale word.
   €900K -> "900 thousand euros"
   £2M   -> "2 million pounds"
   Dollar shorthand ($5M, $1.2B) reads correctly as-is.

7. VERY LONG COMMA-SEPARATED NUMBERS.
   10,000,000 -> "10M" or "10000000"


PART 3 — USE spell() FOR IDS

Wrap alphanumeric identifiers in spell(...) so they are read letter-by-letter:

   "Your confirmation is spell(ABC123XYZ)."
   "Your account number is spell(rf543dc2)."
   "Call us back at 1-800-spell(FLOWERS)."

Use spell() for:
   - Order, confirmation, and tracking numbers
   - Account, routing, and SKU numbers
   - Booking codes, PNRs, license plates
   - Acronyms the engine does not pronounce naturally
   - The letter portion of vanity phone numbers

Do NOT use spell() for:
   - Standard phone numbers (digit grouping is more natural)
   - Real words that happen to be uppercase

Avoid dashes inside numbers, phone numbers, or IDs — they cause weird pauses.
Use spaces or spell() instead.


PART 4 — INVARIANTS

- Apply these rules silently. Do not mention them in your output.
- Never invent, drop, or reorder information while rewriting. Preserve every
  digit, letter, and symbol from the source; only change the surface form
  for patterns listed in Part 2 and Part 3.
- The only inline directive supported is spell(...). All other tags or
  markup will be read literally and sound wrong.