Post-processing
Raw transcripts are accurate but messy: ums, false starts, "like, I mean" trailing off. Post-processing runs each transcript through an LLM and a prompt template, so the final pasted text reads the way you'd write it.
When it runs
Always, if you bound a shortcut to Toggle with post-processing.
Per-session, if you press the post-processing shortcut on this dictation but normally use the plain toggle.
Never, if you turn post-processing off in Settings → Dictation.
Pick a template
Templates live in Settings → Post-processing → Templates. Hyprcore ships with a few defaults you can use as-is or duplicate and edit:
Cleanup. Remove filler words and fix obvious typos. Closest to "what you actually said" without the ums.
Professional tone. Cleanup plus a more polished voice. Good for emails, customer messages, and Slack to coworkers.
Casual tone. Cleanup, but conversational. Good for personal messages, journal entries.
Bullet points. Convert a stream-of-thought into a tight bulleted list. Good for meeting prep and quick notes.
Each template has:
A system prompt that defines tone and rules.
A user prompt template that wraps your raw transcript with
{{TRANSCRIPT}}substitution.A target LLM provider and model.
You can write your own templates—anything from "translate to Spanish" to "rewrite as a JIRA ticket" to "respond as if I'm dictating Markdown."
Pick an LLM provider
Configure providers in Settings → Models → Post-processing.
Provider | Where it runs | Notes |
|---|---|---|
Hyprcore Cloud | Managed via Hyprcore (uses OpenRouter) | Easiest. Pay with tokens from your plan. |
OpenAI | OpenAI's API | GPT-4o or GPT-4o-mini. Bring your own key. |
Anthropic | Anthropic's API | Bring your own API key. Sonnet for quality, Haiku for speed. |
Z.AI | Z.AI's API | GLM models. Bring your own key. |
OpenRouter | OpenRouter's API | One key, many models (including Gemini). Bring your own key. |
Groq | Groq's API | Fast and cheap. Llama Scout / Maverick. Bring your own key. |
Cerebras | Cerebras's API | Fast inference for Llama models. Bring your own key. |
Ollama | Your own machine | Fully local, fully private. Install Ollama separately and point Hyprcore at |
LM Studio | Your own machine | Same idea as Ollama, point Hyprcore at the LM Studio server. |
Custom | Any OpenAI-compatible endpoint | Point Hyprcore at your own URL and model. |
If you use Hyprcore Cloud, the cost is billed in tokens from your plan—see AI tokens.
Performance
Post-processing adds a second or two to dictation. If you're chatting in real time and that lag is annoying, leave the default toggle on plain transcription and reserve the post-processing shortcut for longer drafts.
Tips for good prompts
Lead with what the output should be, not what the input is. "Output: a polished email opening" beats "the user dictated some text."
Tell the model what to keep verbatim—names, code, URLs, technical terms.
For language-switching templates, add an explicit "do not translate proper nouns" instruction.
Use Cleanup as a baseline and stack stylistic rules on top, rather than writing a giant single template.
Was this article helpful?