Prose-level immersion patterns for narrative fiction. Use when writing or revising prose: the sentence-level and paragraph-level craft that pulls readers into the story. Project-specific voice comes from style files passed alongside this skill.
Adversarial reading methodology for narrative fiction: find what doesn't work, not confirm what does. Focus-area driven with dedicated resources per area. Use when reviewing drafts, evaluating prose quality, or assessing changes at any stage.
Arc structure, narrative design, and pacing at multiple scales: saga, arc, chapter, scene. Use when structuring story at any level, planning arcs, designing chapter outlines, or evaluating whether narrative structure serves the story's goals.
Team composition for writing workflows: which agents to spawn, how many, what focus areas to assign, and how to scale effort. Use when composing critic panels, dispatching researchers, staffing draft/revise loops, or setting up brainstorm fan-outs.
Story brainstorming capture: minimal notes that preserve creative freedom. Use when exploring narrative ideas, discussing characters, planning chapters, or thinking through story possibilities.
Logging and referencing writing issues: craft problems, tics, inconsistencies, and structural concerns found during analysis, critique, or review. Use when an agent identifies something worth tracking beyond a single critique report: repeated tics across chapters, inconsistencies that affect multiple scenes, structural problems that need the author's attention, or patterns that should be fixed in revision.
| name | llm-writing |
| type | guardrail |
| description | Load when producing any written artifact for humans. |
LLM training creates default writing behaviors that show up regardless of domain. Recognizing them is most of the battle: once you notice the pull, you can judge whether to resist it or let it through for this particular piece.
Generating fluent text that fills structural expectations without anchoring to purpose. Summarizing with labels instead of explaining how things work. Stating conclusions without showing evidence. Applying the same structure to every document. Smoothing over genuine uncertainty. Encoding corrections as absolute prohibitions. Defining things by what they aren't. Filling space with transitions that restate what was just said.
These aren't always wrong. The failure is when they happen by default rather than by choice.
The most common structural tells come from writing as if responding to a user when producing a document for a reader who wasn't in the conversation. "It's not X; it's Y" corrects a misconception nobody has. "Let's break this down" addresses a question nobody asked. "Here's the thing" builds suspense for a reader who just wants the information. Fractal summaries recap a conversation that didn't happen.
When writing a document, write for the reader. They have no context from the conversation that prompted the document.