writing-staffing

Writing Staffing

Each delegated agent loads its core skills from its definition. This skill teaches which extra skills to attach, files to pass, and resources to reference.

Model Selection

Agent defaults are usually correct. When overriding or selecting models for additional lanes, match the model to the work: creative taste and high-stakes critique need judgment; outlining and cross-document checks need structured synthesis; bulk gathering and mechanical checks can use a cheaper model when later synthesis will catch misses.

For model-diverse fan-out, choose two available model families with different blind spots. Check the live model catalog and model-specific prompting guidance at dispatch time rather than assuming which models or routes are available.

Fan-out passes the same question and files to different model families for independent judgment. Reserve it for high-stakes calls where model diversity can reveal different blind spots. Parallel lanes use different prompts or focus areas; use each agent's default model unless a lane needs a capability it lacks.

Dispatch Reference

@writer

Attach /character-sim for voice fidelity and /shared-dao for project vocabulary.

Name the production mode from /creative-writing-modes → resources/prose-modes.md (fresh draft, revision, bridge, alternate take, line polish). Point to /creative-writing-craft → resources/prose-writing.md or resources/scene-construction.md when relevant. Pass the style and character-state files for everyone who appears, plus the continuity anchors.

One writer per scene — voice consistency degrades when multiple writers handle adjacent content.

@critic

Attach /creative-writing-craft for prose or voice focus and /shared-dao for vocabulary checks.

Assign a focus area: structure, character, voice, prose, or continuity. Pass the style file for voice critique and character-state files for character or continuity focus.

Run different focus areas as parallel lanes. Scale to stakes: 1–2 for low-stakes, 3 for standard chapters, 4–5 for pivotal scenes with duplicated coverage on the critical dimension.

For a pivotal scene or disputed judgment, fan out the same critical dimension across two strong model families, then synthesize the disagreement.

@editor

Name the edit level: editorial review, developmental, line edit, copyedit, or proofreading. Point to /story-review → resources/editorial-review.md for a holistic pass, or the specific edit-level resource.

Use when the draft needs a priority order across concerns. For depth on one dimension, use @critic.

@continuity-checker

Pass the draft plus canon, timeline, character-state, and vocabulary files. This agent is more expensive than a critic with continuity focus because it reads broadly across the project. Use the critic for routine checks and the continuity checker for deep cross-project validation.

@brainstormer

Attach /character-sim for character arcs and /creative-research for real-world grounding.

Run parallel lanes on different angles, not the same angle. Three perspectives beat five instances of one.

@outliner

Outlining starts after direction is chosen — use @brainstormer first. The outliner's output feeds the writer.

@style-creator

Pass sample chapters or existing style files. Point to /creative-writing-craft → resources/style-analysis.md.

@reader-sim

Attach /character-sim when the reader persona is a specific character type. Specify the reader persona and knowledge boundary: what this reader has already read. Pass the draft.

Run after the write/critique loop converges, before presenting to the author. A scene can be technically clean and leave a reader cold.

@character-sim

Pass character-state and style files. Specify the scenario or relationship to explore. Use one parallel lane per character or perspective for independent exploration; use one shared simulation when testing their interaction.

@web-researcher

Specify the question, story context, and what the story currently assumes so the researcher can flag contradictions.

@kb-lead

Attach /story-memory for fiction-specific fact categories and artifact layout. Delegate after the triggering event settles: a chapter is finalized, a brainstorm concludes, or the author makes a decision.

Effort Scaling

Scale critic coverage to stakes. Knowledge maintenance waits until direction or chapters settle. Reader simulation runs after the write/critique loop converges.