AI Voice Inference tool
The AI Voice Inference tool reads a body of content and proposes the voice
profile it is already written in, the inverse of voice-check, which holds
content to a profile that exists. It is the onboarding move: a project that has
been writing in a consistent register for years has a profile, and this reads
it off the corpus rather than asking someone to write it down.
Inference is corpus-level rather than per-block. Every part passes through unchanged while the tool accumulates each block's source text, and one inference runs when the input stream ends. The corpus is truncated on a rune boundary before it is sent, so a large document set costs one bounded request. An empty stream leaves nothing to infer from, so the tool reports success and yields no draft.
The result is a draft in the ordinary voice-profile shape (tone, style,
vocabulary and before/after examples), read from Draft() after the flow
completes, together with an evidence sidecar carrying the model's confidence
and a short source note for each top-level field. The evidence stays beside the
profile rather than inside it, so the draft round-trips through the same
profile.yaml an author edits by hand. Treat the output as a draft: it is a
proposal to review and commit, not a decided profile.
Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
apiKey | string | API key for the AI provider | |
domain | string | Subject domain hint for the analysis (e.g. medical legal technology) | |
maxExamples | integer | 3 | Maximum before/after examples to infer |
model | string | AI model name | |
profileName | string | Name for the inferred draft profile | |
provider | string | anthropic | AI provider |
Configure it live
Examples
Draft a profile from a documentation set
Name the draft and hint at the subject domain; the corpus is whatever the flow reads.
provider: anthropic
profileName: House Style
domain: developer tools
Processing notes
Every part passes through unchanged; the tool annotates nothing on the content.
The corpus is each block's source text, accumulated over the whole stream; inference runs once, when the stream ends.
An empty or whitespace-only corpus yields no draft rather than an error.
Limitations
Requires provider credentials and makes one external API call per run.
The corpus is truncated before it is sent, so a very large content set is inferred from its beginning.
The output is a draft to review; confidence is the model's own estimate, not a measurement.
← Back to the Tool Reference