AI Enhancement
The raw transcript is just the beginning. AI Enhancement passes it through a language model with a prompt of your choice — to fix grammar, reformat for email, condense rambling, prettify code, translate, summarise, or anything else you can describe in a system prompt.
It's optional. Neuron Flame works perfectly fine without it. Enhancement is a layer on top — turn it on only when you want polish.
Supported providers
| Provider | Models | Where it runs |
|---|---|---|
| Google Gemini | Flash, Pro, Flash Lite | Google cloud |
| OpenAI | GPT-4o, GPT-4o-mini, etc. | OpenAI cloud |
| Anthropic | Claude 3.5 Sonnet, Haiku | Anthropic cloud |
| Ollama | Llama, Mistral, Qwen, anything local | Your Mac |
| Groq | Llama 3.3, Mixtral | Groq cloud (fast) |
| Custom OpenAI-compatible | Any URL + model name | Wherever you point it |
| Local CLI | Any executable | Your Mac |
Setting up
- Open the AI Models tab.
- Choose a provider.
- Paste your API key (Gemini and Anthropic offer generous free tiers).
- Pick a model from the dropdown.
- Toggle Enable AI Enhancement on.
That's it. Every transcription now flows through the model before being pasted.
Prompts
Neuron Flame ships with a curated set of prompts:
- Default — light cleanup. Fixes grammar, removes filler words, keeps your voice.
- Email — formats as a professional email with greeting and sign-off.
- Chat — short, casual, no formality.
- Notes — bullet-formatted summary.
- Code — preserves identifiers, fixes only natural-language portions.
You can edit these or create your own. Each prompt has a system message and an optional user-message template — both support variables like {{transcript}}, {{clipboard}}, and {{window_context}}.
Context-aware enhancement
Two optional toggles make the model smarter:
- Clipboard context — the model sees what's currently on your clipboard. Useful for "make this paragraph more formal" or "summarise the article I just copied".
- Window context — the model sees the title and URL of the frontmost window. Useful for "reply to this email" or "comment on this PR".
Both are off by default. Both are processed locally and sent only to the model you've configured. Neither is ever written to disk.
What happens if it fails?
If the LLM call fails — rate limit, network blip, invalid key, anything — Neuron Flame falls back to the raw transcript. You always get something pasted; you never get an error string in your document. The History tab shows the raw transcript only (the failed enhancement is not stored).
Custom prompts
The prompt editor uses Markdown for the system message. Variables you can reference:
{{transcript}}— the raw text from Whisper/Parakeet (with your custom dictionary applied).{{clipboard}}— current clipboard text (only if Clipboard Context is enabled).{{window_context}}— frontmost app + window title + URL (only if Window Context is enabled).{{custom_vocabulary}}— your custom dictionary entries (handy for prompts that need to know specific spellings).
The wrapper artifacts (<TRANSCRIPT> tags etc.) are stripped from the model's reply automatically — even if it tries to echo them back, they won't end up in your text.
Privacy notes
- If you use a cloud LLM, your transcript leaves your Mac and goes to that provider. That's the trade-off.
- If you use Ollama or a local CLI, nothing leaves your Mac — you get the polish without the privacy cost.
- API keys are stored in your macOS Keychain. Neuron Flame never logs or transmits them.
See the full Privacy doc for the complete data-flow diagram.