Voice and style profile

Returns typed style records with their source-note evidence. Use style_profile (POST /v1/style/profile) when an AI tool needs to write in this person's voice (résumé bullets, cover letters, bios, talks, email drafts) without prompt-stuffing the source corpus.

Style records are real graph entities with provenance, confidence, and edges back to the notes they were learned from. They are not summarised paragraphs of advice; they are structured rules a writing agent can reason over.

• Three categories: style_observation, style_preference, vocabulary_rule.
• Every record carries the evidence-note ids it was extracted from.
• Optional scope filter narrows to a specific writing context (e.g. technical-writing, formal-writing).
• Optional sourceRunId filter narrows to records extracted from a single ingest run, useful for showing what a fresh thesis or essay actually contributed to the voice.

#Example

{ "scope": "technical-writing", "limit": 8 }

Every response carries four parts:

#Observations

Descriptive patterns the distiller noticed. Each carries an optional consumerGuidance field so an agent can act on it directly.

[
  {
    "id": "a01f…",
    "type": "style_observation",
    "category": "structural",
    "scope": ["technical-writing"],
    "pattern": "Prefers concrete nouns over abstractions; opens paragraphs with the subject acting.",
    "consumerGuidance": "Start sentences with the actor; avoid passive scaffolding.",
    "confidenceKind": "asserted",
    "confidenceScore": 0.82,
    "evidenceNoteIds": ["9b2c…", "11ee…"],
    "rationale": "Across five distilled paragraphs the subject-first pattern held."
  }
]

#Preferences

Prescriptive rules: what the writer wants downstream output to do or avoid.

[
  {
    "id": "b22d…",
    "type": "style_preference",
    "category": "framing",
    "scope": ["technical-writing", "formal-writing"],
    "preference": "Frame the Northstar work as cross-functional delivery, not just a product launch.",
    "strictness": "strong",
    "confidenceKind": "user-asserted",
    "confidenceScore": 1.0,
    "evidenceNoteIds": ["7a8b…"],
    "rationale": "Direct user-context assertion."
  }
]

#Vocabulary rules

Specific terms, casing, or substitutions the writer cares about.

[
  {
    "id": "c44e…",
    "type": "vocabulary_rule",
    "category": "casing",
    "scope": ["technical-writing"],
    "term": "Northstar",
    "rule": "Capitalize as a proper noun; never split into two words.",
    "strictness": "strong",
    "confidenceScore": 0.9,
    "evidenceNoteIds": ["9b2c…"]
  }
]

#Evidence notes

For every style record above, the evidence-note excerpts behind it, so an agent can quote real voice rather than paraphrasing rules.

[
  {
    "id": "9b2c…",
    "sourceRunId": "f1a3…",
    "sourceKind": "blog-post",
    "snippet": "We did not pitch Northstar as a product launch. We pitched it as a cross-functional delivery effort that unified Product, Design, and Operations …",
    "styleDescription": "Technical, declarative, foregrounds the subject."
  }
]

#Filtering

Filters are non-destructive: if scope matches no records, you get empty arrays plus the evidence notes for any matched records (which will also be empty in that case).

#Building a style guide

A writing agent can compose observations + preferences + vocabularyRules into a short style guide, then back-cite using evidenceNoteIds. The local CLI's marrow style eval-plan and marrow style eval-run build blind A/B writer prompts and a source-aware judge prompt on top of this same data. See the CLI integration.

#What you get back

#Scope

Requires an API key with scope style. See the Style Profile API Reference for the full schema.