How to use the story-deslop skill to de-AI Chinese web novels

What an agent skill is and what story-deslop does

An agent skill is a folder of instructions and reference files that an AI coding or writing agent loads when a task matches it. story-deslop is one such skill, written for Chinese web fiction.

When it runs on a chapter, it reads the text in three layers. Expression covers wording, sentence patterns and repeated templates. Narrative focus covers where the writing spends and saves its detail. Story logic covers whether actions, motives, consequences and what each character knows actually hold together.

It fixes problems it can fix directly, such as stiff or formulaic wording. Purely story-level gaps are reported as options for the author to approve rather than rewritten on its own. It is not an AI-detection tool and does not promise any detection score.

• Keeps the author's style and the plot already decided.
• Does not add new plot facts, motives or outcomes without approval.
• Reports unresolved story gaps with the original location and the smallest design choice needed.

Install story-deslop from the ZIP

Download the story-deslop ZIP from this directory page and unzip it. You can install it for yourself across projects or for a single project.

For a personal install, put the unzipped skill folder in ~/.claude/skills/. For a project-only install, put it in the project's .claude/skills/ folder. Start a new session afterwards so the agent picks it up.

If you use the Claude apps instead of a local setup, upload the skill under Settings, then Capabilities, then Skills.

• Personal: ~/.claude/skills/
• Project only: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills

Trigger the skill with natural requests

You can invoke the skill with the slash command or by asking in plain Chinese. The skill is triggered by /story-deslop, /去AI味, and requests such as 去AI味, 这篇太AI了 or 网文去AI味.

Say what scope you want. If you only want a check, say so, because a detection-only request produces a diagnosis and leaves the text unchanged. If you say only 去AI味 or 这篇太AI了, the skill treats it as a revision request: it edits expression and narrative focus directly and proposes options for story gaps.

• Diagnosis only: ask it to check or annotate, not to change the text.
• Full revision: 去AI味 or 这篇太AI了 on a chapter file.
• Limited scope: name a paragraph or a problem type, such as dialogue or pacing, and it will limit the review to that.

Run the checks on a manuscript

For a local Chinese manuscript, the skill first runs a deterministic check script that only reports. Its hits are leads for the agent to examine in context, not verdicts, and zero hits do not replace the three-layer read. After editing, it runs the check again and then a degeneration check that looks for repeated text, truncation, placeholders and leaked engineering words.

Run from the skill's root folder, with manuscript paths as absolute paths. The skill's own instructions use this form for the degeneration check.

• Before editing, the expression check script reports blocking and advisory items. Blocking items should be revised or kept only through the style whitelist.
• After editing, both scripts run again. Exit code 1 means blocking items remain; exit code 2 means a command or file error.
• Long projects with a 正文/ folder and 追踪/ folder: plot-changing edits are handed to /story-long-write so the original is backed up and tracking is updated.

Know what the skill needs and touches

You need Node.js to run the bundled scripts, and a local Chinese manuscript file such as 正文.md or chapter files. The companion narrative-writer agent is optional: in Claude Code or similar hosts the skill can use it for the editing pass, and if it is unavailable the skill edits directly.

The skill runs scripts. It does not use the network, according to its listing. Before changing a file it copies the original into the book's .story/work/去味原稿/ folder so you can restore it. It loads author preferences only from the calling workflow and does not read or write a separate memory file when run alone.

• Requires: Node.js, a local manuscript, optional narrative-writer agent.
• Runs scripts: yes. Uses network: no.
• Backup of the original goes to .story/work/去味原稿/ before editing.
• A chapter marked with the skip line can still be processed when you ask for de-AI editing; the skill reports whether it is still treated as skipped.

Tips for getting useful results

Give the skill one chapter or a clearly bounded set of chapters at a time. It works through each chapter in order, reading the text and related earlier material as needed.

Read the report before accepting changes. Blocking items that remain are listed for you. Story-level gaps need your decision, and the skill will not add a plot condition to fill one.

If your project is long and a fix would change events, relationships or outcomes, take the diagnosis to /story-long-write rather than asking this skill to rewrite the plot.

• Name the scope: a paragraph, a chapter, or a problem type.
• Keep the source file; the skill's backup is a second safety net.
• Expect the report to separate fixed issues from items that need your design decision.