Style Research Agent

The Style Research Agent runs a focused LLM research pass before outlining begins, extracting concrete, labelled style attributes from the inspired_by field and persisting them in the ProjectKnowledgeBase.

Overview

When a user sets inspired_by (e.g. "Cormac McCarthy", "The Road", "Brandon Sanderson's Stormlight Archive"), this agent converts that raw string into a structured StyleProfile — eight specific, actionable style constraints that every downstream agent can act on.

Without this agent, style knowledge is implicit and inconsistent across LLM providers. With it, there is a deterministic record of exactly which style attributes were applied.

When It Runs

StyleResearchAgent is called once by the Project Manager inside generate_outline(), before the Outliner runs. It is a no-op when:

• inspired_by is empty or whitespace
• pkb.style_profile is already set (resumed run — no extra LLM call)

StyleProfile Fields

Field	Description	Example
point_of_view	Narrative perspective	"Close third-person limited, single protagonist"
sentence_length	Sentence rhythm	"Short declarative sentences, rarely exceeding 15 words"
pacing	Story tempo	"Slow burn with sudden violent accelerations"
dialogue_style	How characters speak	"Sparse, unattributed, subtext-heavy"
prose_tone	Overall voice	"Bleak, lyrical, biblical cadence"
structural_patterns	Chapter/section structure	"Non-linear timeline, chapter-length vignettes"
thematic_preoccupations	Recurring themes	"Violence, grace, fate, the American West"
what_to_avoid	Anti-patterns	"Purple prose, extended exposition, comic relief"

Pipeline Integration

Once populated, pkb.style_profile is injected as a STYLE REFERENCE block into:

• OutlinerAgent — both the chapter outline prompt and individual scene outline prompts
• ChapterWriterAgent — every scene prompt via SCENE_PROMPT.format(style_profile_block=...) and the prefix_parts prepend block

When inspired_by is absent, style_profile_block is "" and no STYLE REFERENCE block appears.

Persistence

StyleProfile is a Pydantic model stored as a nested object in project_data.json. It survives save/load round-trips via the existing PKB serialisation path — no migration needed.

Error Handling

• Malformed JSON: A regex extraction fallback attempts to find a {...} block in the LLM response before giving up.
• LLM failure: Logs an exception and continues. The pipeline never stops because of a missing style profile.
• Partial JSON keys: Extra keys from the LLM are silently ignored. Missing keys default to "".

Usage

The agent is invoked automatically. No standalone CLI command is needed. To trigger it, set inspired_by during Advanced Mode setup or in an expert config file:

inspired_by: "Cormac McCarthy"

Integration

Works with:

• Project Manager — registers and calls StyleResearchAgent before outlining
• Outliner Agent — receives STYLE REFERENCE block in chapter and scene outline prompts
• Chapter Writer — receives STYLE REFERENCE in every scene prompt and prefix_parts