Claude Skills for Novel Writing: How to Build One
You have a paragraph you paste. Maybe it's 600 words, maybe it's 1,200. It says your third-person limited stays locked to one head per scene, that your dockhand speaks in fragments and your archivist never does, that you want scenes rendered rather than summarized, and that if you get one more character who lets out a breath she didn't know she was holding, you're closing the laptop. You paste it at the top of every session. Last Tuesday you forgot the fragment rule, and chapter 19 came back sounding like a press release.
Claude Skills fix that paste. Not the bigger problem — we'll get to what they can't touch — but that one, completely and permanently.
What a Claude Skill actually is
A Skill is a folder containing a SKILL.md file: YAML frontmatter with two required fields, name and description, followed by a markdown body of instructions. That's the entire format. Anthropic's documentation defines Skills as "modular capabilities that extend Claude's functionality", packaging "instructions, metadata, and optional resources (scripts, templates) that Claude uses automatically when relevant."
Don't let the word "skill" imply more than it is. It isn't a model, a fine-tune, a paid add-on, or a marketplace purchase. It's a text file in a folder, and you can write a useful one in about ten minutes.
The part worth understanding is the loading mechanism, which Anthropic calls progressive disclosure. It runs in three stages. At startup, Claude loads only the name and description from every installed Skill — roughly 100 tokens each, by Anthropic's own accounting. When your request matches a description, Claude reads the SKILL.md body into context; the docs target under 5k tokens there. Everything else in the folder — reference files, sample scenes, templates — costs nothing until Claude actually opens it.
That third tier is the one novelists should care about. You can bundle a forty-page style guide, six scenes written in your voice, and every phrase you've ever banned, and none of it competes for attention until the moment it's needed. As Anthropic's engineering team put it, the model doesn't "need to read the entirety of a skill into their context window," which makes the amount of bundled material "effectively unbounded."
Why this matters differently for novelists
Because your context problem has two halves, and a Skill solves exactly one of them.
Half one is procedure: how you want the work done. POV rules, prose allergies, beat structure, the fact that your chapters run 2,800 words and not 4,000. This is stable across an entire book and probably across your career.
Half two is state: what happens inside the book. Who carries the scar, who knows about the treaty, which week the coronation falls in. That changes every chapter, and it's the half that quietly wrecks AI-assisted novels — the model confidently filling gaps it can't see.
A Skill remembers how you write. It does not remember what you wrote. Every writer who tries to cram a story bible into a SKILL.md discovers this around chapter twelve, usually when a minor character's eye color changes mid-scene despite the file sitting right there on disk.
The distinction doubles as a sorting filter. If an instruction would still be true tomorrow if you abandoned this book and started a completely different one, it belongs in a Skill. If it's a fact about this book, it belongs in a codex retrieved per scene — the approach NovelMage's Codex takes, pulling only the entries a given scene touches rather than asking you to re-paste your world every session.
Building your first Skill, step by step
Here's the build, in the order that actually works.
1. Find the thing you keep re-pasting. Anthropic's own authoring guidance recommends working through a task normally first and noticing "what information you repeatedly provide." For most novelists that's a voice brief, a list of banned constructions, or a chapter-shape template. Don't invent a Skill for a workflow you don't already have.
2. Make the folder. In Claude Code, personal Skills live at ~/.claude/skills/<skill-name>/SKILL.md and project Skills at .claude/skills/<skill-name>/SKILL.md — the docs cover both, plus plugin-scoped Skills. On claude.ai you upload a Skill as a zip through Settings > Features, available on Pro, Max, Team and Enterprise plans with code execution enabled. No manuscript required either way; the folder is just text.
3. Write the frontmatter, and take the description seriously. This is the field Claude matches your request against, so it must say both what the Skill does and when to use it. Constraints: name maxes out at 64 characters, lowercase letters, numbers and hyphens only, and cannot contain the reserved words "anthropic" or "claude." description maxes out at 1,024 characters and must be written in third person — Anthropic warns that first- or second-person phrasing "can cause discovery problems," because the text gets injected straight into the system prompt.
---
name: line-editing-drafts
description: Line-edits fiction drafts against a fixed voice spec — cuts filter verbs, preserves fragment-heavy dialogue for working-class characters, flags head-hops in third limited. Use when the user asks for a line edit, prose polish, or voice check on a chapter or scene.
---
Compare that to description: helps with writing. The first one gets picked correctly out of thirty installed Skills; the second one never fires, or fires on your grocery list.
4. Write the body at the right level of freedom. Anthropic frames this as matching specificity to fragility: high freedom (plain instructions) when multiple approaches are valid, low freedom (exact scripts) when a sequence is brittle. Craft is almost entirely high-freedom territory. "Cut filter verbs — she saw, he felt, she noticed — unless the filtering is the point of the sentence" is a rule; "rewrite every third sentence" is a recipe for damage.
5. Test it against a real chapter, not a toy one. The best-practices guide is blunt about this: build evaluations before writing extensive documentation, run Claude on representative tasks without the Skill first to establish a baseline, then measure. For fiction that means taking a chapter you already line-edited by hand and seeing whether the Skill reproduces your calls.
6. Split it when it grows. Keep the SKILL.md body under 500 lines, and move overflow into sibling files linked directly from it. One caveat from the docs that bites people: keep references one level deep. If SKILL.md points at voice.md which points at dialogue.md, Claude may only preview the nested file rather than read it whole.
A worked example
Marisol is 91,000 words into a three-POV second-world fantasy. Her pasted voice brief had grown to 740 words over eight months — POV rules, a banned-phrase list, notes on how each POV character's interiority differs, and a reminder that her merchant-prince narrator uses subordinate clauses the way other people use punctuation.
She split it into a Skill in an afternoon. The SKILL.md came out at 118 lines: frontmatter, the four rules that apply to every scene, and pointers to three reference files — voice-per-pov.md (about 900 words), banned-constructions.md (a flat list of 62 entries), and scene-shape.md (her 2,800-word chapter template with two annotated examples).
The measurable change wasn't quality, it was consistency. Before, she pasted the full brief maybe six sessions out of ten, and pasted a shortened version the rest of the time because 740 words at the top of every chat got tedious. Now it's ten out of ten, and the banned-phrase list stopped drifting because it lives in a file she edits rather than a paragraph she retypes. Her first three test chapters came back with zero head-hops, down from a handful per chapter — because the POV rule stopped being the thing she forgot to include when she was tired.
The part she didn't expect: the reference split meant the model only pulls voice-per-pov.md when she's drafting, not when she's asking about structure. Three files, loaded independently, none of them crowding the others.
Three Skills worth building before your next chapter
The voice pass. Everything above. This is the highest-leverage Skill for a working novelist because voice is where AI output goes generic first, and it's a problem of keeping your own register intact rather than of raw model quality.
The continuity check. Feed it a chapter and your codex entries and have it flag contradictions. One rule to build in from the start: instruct it to accept a list of intended contradictions first. A planted secret is structurally identical to a continuity error — your best foreshadowing will get flagged as a bug on every single run unless you declare it.
The synopsis and query pass. These are a different register entirely: present tense, spoil everything, 500–750 words, no blurb voice. Left to itself the model reliably writes back-cover copy instead, because that's what "summarize my novel" looks like in training data. A Skill that spells out the form once saves you the argument every time you query.
Where Skills stop
Some real limits, all documented, all worth knowing before you build a workflow on top of them.
Skills don't sync across surfaces. A Skill uploaded to claude.ai isn't available through the API, and Claude Code's filesystem Skills are separate from both. You manage each surface yourself.
On the Claude API, Skills run sandboxed with no network access and no runtime package installation — they need the code execution tool and a beta header, which is why a desktop app that calls Claude with your own API key isn't where a Skill lives. Skills are an Anthropic-surface feature, not something that travels with the model into every client.
Only install Skills you trust. Anthropic's security note is direct: a malicious Skill "can direct Claude to invoke tools or execute code in ways that don't match the Skill's stated purpose." Treat a downloaded Skill like installing software, and read the whole folder before you run it.
And if you draft one novella a year in a single long chat and never repeat yourself, honestly skip all of this — the setup costs more than the paste does.
The open standard part
Skills aren't a Claude-only format, which matters if you've been burned by tools that hold your work hostage. The format was originally developed by Anthropic and released as an open standard, and the client list now includes Cursor, GitHub Copilot, VS Code, Gemini CLI, OpenCode, Codex and a long tail of others. Anthropic also publishes example Skills in a public repository — many under Apache 2.0, with the production document Skills included as source-available reference material for "more complex skills that are actively used in a production AI application."
Practically: your voice spec is a plain-text folder you own, version-control, and carry to another agent if you change your mind. That's the same reason writers end up preferring a desktop editor to a browser tab — the local-model setups worth running keep an unpublished manuscript on your own disk rather than in someone's retention policy, and Anthropic notes separately that Agent Skills are not covered by zero-data-retention arrangements.
If you want the drafting environment to work the same way — your manuscript on your machine, your codex injected per scene, your choice of Claude, GPT, Gemini or a local model through Ollama or LM Studio — NovelMage is $99.99 once for up to three devices, future updates included, with a 7-day trial that doesn't ask for a card.
Frequently Asked Questions
Do I need Claude Code to use Skills for novel writing?
No. claude.ai supports custom Skills uploaded as a zip through Settings > Features, on Pro, Max, Team and Enterprise plans with code execution enabled. Claude Code is the more convenient path if you're comfortable with a folder on disk, since Skills there are just directories at ~/.claude/skills/ that you can edit in place and version-control.
Can a Skill hold my entire story bible?
It can hold the files, but it's the wrong shape for the job. Progressive disclosure loads a whole file once Claude decides to read it, so a 12,000-word bible arrives all at once and every irrelevant entry competes with the one that matters. Story state wants per-entry retrieval scoped to the scene you're writing. Keep procedure in the Skill and facts in a codex.
Which Claude model should I pair a writing Skill with?
Anthropic recommends testing a Skill against every model you plan to use it with, because what reads as over-explaining to a larger model may be exactly the guidance a smaller one needs. For fiction specifically, the model-by-model breakdown for creative writing is the better starting point than a generic benchmark.
How long should a SKILL.md be?
Under 500 lines for the body, per Anthropic's best-practices guide, with the loaded instructions targeting under 5k tokens. Past that, split into reference files linked one level deep from SKILL.md. Conciseness genuinely matters here: the guide's framing is that "the context window is a public good," and every token your Skill spends is a token your chapter doesn't get.