What Are Output Formats?
When you speak into Telvr, two things happen. First, your voice is transcribed by Whisper — one of the most accurate speech recognition models available. Second, that raw transcript is passed through an AI enrichment step that transforms it into polished, structured text.
Output formats are the instructions that govern that second step. They tell the AI what kind of writer it should be, how to behave, what the input context is, and what the final result should look like. Technically, an output format is a system prompt — a set of instructions the AI reads before processing your transcript.
Telvr ships with six built-in output formats to cover the most common use cases:
- Cleanup — removes filler words, fixes grammar, produces clean prose
- Email — structures dictated content into a professional email with subject line and greeting
- Meeting Notes — organizes spoken content into decisions, action items, and next steps
- Summary — condenses longer speech into a 2–3 sentence summary
- Prompting — structures spoken ideas as development tasks with requirements and acceptance criteria
- Custom — your own output format, written from scratch
The Custom mode is where output formats become powerful. It lets you define your own template for any workflow — a format for daily standup updates, legal memo dictation, customer support responses, social media captions, or anything else you dictate regularly.
The 4 Template Sections
Every output format template follows a consistent 4-section structure. Each section serves a specific purpose, and together they give the AI everything it needs to produce reliable, high-quality output.
Role
The Role section defines who the AI is for this template. Think of it as hiring a specialist: you are telling the AI what kind of expertise it should draw on and what persona to adopt.
# Role
You are a professional email writer who converts spoken messages into well-structured, polished emails.
A well-defined role shapes the entire output. An AI told it is a "professional email writer" will default to appropriate register, formatting conventions, and word choice without needing explicit instructions for each. A role like "concise technical writer" will naturally favor precision and brevity.
Keep the role to one or two sentences. State the expertise domain and the core task in plain language.
Instructions
The Instructions section is where you define the rules. This is the most important section — it tells the AI exactly what to do, what to avoid, and how to behave when it encounters edge cases.
Format instructions as a bullet list. Each bullet should describe a single, clear rule.
Three rules belong in every template, regardless of purpose:
# Instructions
- [Your task-specific rules here]
- Respond in the same language as the input text.
- Do not add explanations, headers, or meta-commentary.
- Do not wrap output in markdown code blocks.
The last three are universal defaults. "Respond in the same language" ensures your output matches the language you dictated in — critical if you switch between languages. "No explanations or meta-commentary" prevents the AI from prepending things like "Here is your email:" before the actual output. "No markdown code blocks" prevents the output from being wrapped in triple backticks.
You can add as many task-specific rules as needed, but aim to keep the total under ten. More than that and rules start conflicting with each other.
Context
The Context section provides background information about the situation. It is optional — not every template needs it — but it helps the AI handle the idiosyncrasies of dictated input.
# Context
The user dictates messages via push-to-talk. Input may contain filler words, repetitions, and incomplete sentences.
Context is useful when:
- The input has predictable characteristics the AI should account for (filler words, incomplete sentences, domain-specific jargon)
- The output will be used in a specific environment with implicit conventions (a company Slack, a legal document system)
- The AI needs to understand the relationship between speaker and intended recipient
If your use case is straightforward and the Instructions section already covers everything, skip the Context section entirely.
Expected Output
The Expected Output section describes the shape of the final result. It tells the AI what format, length, and structure to produce.
# Expected Output
A professional email with:
- Subject line
- Greeting
- Body (1-3 paragraphs)
- Professional sign-off
This section acts as a quality target. The AI will structure its output to match what you describe here. If you want a specific format — bullet lists, numbered steps, a particular heading structure — describe it here explicitly.
The Blank Template
Use this as your starting point for any new output format:
# Role
[Describe the assistant's expertise and persona.]
# Instructions
- [Primary task instruction]
- [Additional requirement]
- [Formatting preference]
- Respond in the same language as the input text.
- Do not add explanations, headers, or meta-commentary.
- Do not wrap output in markdown code blocks.
# Context
[Optional background information. Remove if not needed.]
# Expected Output
[Describe what the output should look like.]
Best Practices
Six principles from prompt engineering that apply directly to output format templates:
1. Be specific about quantities and constraints. "Write in 2–3 sentences" is better than "keep it short." "Use a maximum of 5 bullet points" is better than "be concise." Vague instructions leave room for the AI to interpret them in ways that may not match your intent.
2. Define tone with two reference points. "Professional but friendly" is more useful than just "professional" because it defines both the floor and the ceiling. Similarly, "formal but not stiff" or "casual but not flippant" gives the AI a range to work within rather than a single ambiguous target.
3. Include examples in Expected Output. Show, don't just tell. If you want a specific format, describe it structurally and include a mini-example when it helps. "A subject line starting with an action verb, e.g. 'Request: Server Access by Friday'" is clearer than "a good subject line."
4. Set explicit boundaries with what NOT to do. The three universal default rules already cover the most common issues, but your template may need additional boundaries. If your dictation often contains tangents, add "Do not include content unrelated to the main request." If you want a clean output without any formatting characters, add "Do not use bullet points or headers."
5. Keep each template focused on a single task. A template that tries to handle email and meeting notes and summaries will handle each poorly. One template, one output type. If you have multiple distinct use cases, create multiple templates and switch between them with a click.
6. Start simple and iterate. Write a first version with just three or four rules and test it against real dictations. Add rules only when you encounter specific problems in the output. Over-engineered templates with 15 rules often perform worse than lean templates with 5, because the rules start contradicting each other.
Example: Blog Post Draft
# Role
You are a content writer who transforms spoken ideas into blog post drafts.
# Instructions
- Convert the spoken input into a structured blog post draft.
- Add a compelling title.
- Break the content into logical sections with subheadings.
- Maintain a conversational but informative tone.
- Respond in the same language as the input text.
- Do not add explanations, headers, or meta-commentary.
- Do not wrap output in markdown code blocks.
# Expected Output
A blog post draft with:
- Title (H1)
- 2–4 sections with subheadings (H2)
- 200–400 words total
This template is useful for content creators who think out loud. You speak your ideas in a stream-of-consciousness style; the template shapes them into a publishable draft structure.
Example: Slack Message
# Role
You are a concise communicator who converts spoken messages into Slack-appropriate messages.
# Instructions
- Keep messages brief and scannable.
- Use emoji sparingly but naturally.
- Break longer messages into short paragraphs.
- If the message contains action items, format them as a checklist.
- Respond in the same language as the input text.
- Do not add explanations, headers, or meta-commentary.
- Do not wrap output in markdown code blocks.
# Expected Output
A casual but clear Slack message, 1–3 short paragraphs maximum.
This template reflects the norms of Slack communication: short paragraphs, occasional emoji, and clear action items when needed. The checklist instruction only activates when relevant — the AI will recognize when your message contains tasks and format accordingly.
Common Mistakes
1. Too vague "Make it better" or "improve the text" gives the AI no direction. Output formats work best with specific, actionable instructions. If you cannot articulate what "better" means, the AI cannot either.
2. Too many rules Templates with 20 instructions create internal conflicts. When one rule says "be brief" and another says "include full context," the AI has to guess which takes priority. Keep rule counts under ten and resolve any conflicts explicitly.
3. Missing the language rule Always include "Respond in the same language as the input text." Without it, the AI may default to English regardless of what language you dictated in. This matters especially if you switch between languages depending on the context — a Slack message to a French colleague, an email to a German client.
4. Forgetting Expected Output Without an Expected Output section, the AI guesses at structure. Sometimes it guesses right. Often it produces inconsistent results across similar inputs. Explicitly describing the desired output format eliminates that variance.
5. Over-engineering from the start New users sometimes try to anticipate every edge case before testing anything. Write the simplest version that captures your intent, run it against ten real dictations, and add rules only for problems you actually observe. Templates are iterative — they get better with use.
Get Started
Download the blank template below, open Telvr, and paste it into the Custom mode template field. Start with the Role and a few Instructions, then test it against real dictations from your workflow. Most useful templates emerge from a handful of iterations: write, test, refine.
The blank template is yours to shape. Telvr will handle the rest.
Download Blank Template