writing-editor（写作编辑）

Writing Editor

Edit prose using a top-down workflow with human review at each level. Works for any nonfiction genre — academic papers, reports, memos, essays, blog posts, proposals.

Project Integration

This skill reads from project.yaml when available:

# From project.yaml
paths:
  drafts: drafts/sections/

Project type: This skill works for all project types. Prose editing improves writing regardless of methodology.

File Management

This skill uses git to track progress across phases. Before modifying any output file at a new phase:

1. Stage and commit current state: git add [files] && git commit -m "writing-editor: Pre-editing snapshot"
2. Then proceed with modifications.

Do NOT create version-suffixed copies (e.g., -v2, -final, -working). The git history serves as the version trail.

Workflow: Four Levels with Checkpoints

Work through each level, presenting proposed changes for user approval before moving to the next.

Step 0: Document Protection

Before making any edits:

• Check if the file is in a git repo; if not, offer to git init
• Commit the original state before any edits: git add [file] && git commit -m "writing-editor: Pre-editing snapshot"
• After each level's approved changes, commit with a message like: "Writing editor: Level 1 (Document) complete"

This creates a full revision history the user can diff or revert. The input file IS the output — edit it in place. Do not create a working copy or a separate output file.

Levels with Checkpoints

This ensures the user stays in control and can accept/reject changes at each stage.

Quick Start

/writing-editor

Please edit: /path/to/draft.md

Or with pasted text:

/writing-editor

Here's a draft that sounds too formal: [paste text]

Primary Reference

Use references/merged-guidelines.md as the main editing guide. It consolidates all rules organized by level:

• Level 1: Document (6 rules) - hooks, titles, structure, abstracts, citations, concrete examples
• Level 2: Paragraph (10 rules) - endings, symmetry, triplets, contrast, syntax-logic match, listicle structure, gerund fragments, one-point dilution, signposted conclusions, "despite challenges" formula
• Level 3: Sentence (20 rules) - passive voice, first person, abstract nouns, placeholders, agents, rhetorical questions, "serves as," dramatic countdown, false ranges, invented labels, superficial analyses
• Level 4: Word (9 rules) - throat-clearing, signposts, adverbs, intensifiers, LLM vocabulary, magic adverbs, fancy-talk

Additional References

For deeper context or source-specific guidance:

Core Method: Deletion Test

At every level, apply Becker's test: Remove each word or phrase. If meaning doesn't change, delete it.

Level 1: Document

Before touching sentences, fix:

• Opening hook: Does it grab attention or start with a bland formula?
• Title: Short and unified, or bloated with variables and colons?
• Structure: Do section headings match what the opening promises?
• Abstract/summary: Active voice with humans and claims, or passive hedging? (Skip if genre has no abstract.)
• Citations/references: Do they advance the argument or just signal allegiance? (Adjust for genre — academic papers cite; memos and blog posts may not.)
• Concrete examples: Is each major concept grounded in specifics?

Present document-level changes. Wait for user approval.

Level 2: Paragraph

After document structure is sound:

• Paragraph endings: Do they moralize ("Together, these underscore...") or just stop?
• Symmetry: Do three paragraphs start the same way?
• Triplets: Ornamental lists of three that could be two? (Keep conceptual triplets like "race, class, and gender")
• Over-balanced contrast: "Not X, but Y" that could be one clause?
• Syntax-logic match: Does grammar show which ideas are subordinate?
• Sentence length variation: Do all sentences in a paragraph have similar word counts? Flag uniform-length paragraphs. Good paragraphs mix short (10-15 words) and longer (25-30) sentences with ~10 words stdev.
• Paragraph length: Flag paragraphs exceeding 6 sentences for possible splitting. Median in published sociology is 3 sentences.
• Stacked intensifiers: Does the paragraph keep raising emotional volume sentence after sentence? One strong claim is enough.
• Listicle in disguise: "The first... The second... The third..." dressed up as continuous prose? Either use an actual list or integrate without ordinal labels.
• Gerund fragment litany: Standalone "-ing" fragments used to illustrate a claim? Fold into a single sentence or cut.
• One-point dilution: Does the section restate the same argument in multiple ways without adding anything? Cut redundant restatements.
• Signposted conclusions: "In conclusion," "To sum up," "In summary" — cut these. The reader can feel the ending.
• "Despite its challenges" formula: Acknowledging problems only to immediately dismiss them? Engage with the challenges or don't raise them.

Present paragraph-level changes. Wait for user approval.

Level 3: Sentence

After paragraphs are structured:

• Passive voice: "Data were collected" → "We collected data"
• First person: Use I/we for methods and claims
• Abstract nouns: "The investigation of" → "We investigated"
• Placeholders: "complex relation" → specify the actual relation
• Deictic pronouns: "This shows" → "This finding shows"
• There is/are: "There is evidence" → "Evidence shows"
• Subject-verb distance: Keep within 12 words
• Vivid verbs: Replace weak verbs with specific action
• Dead metaphors: Cut "cutting edge," "shed light on"
• Meta-commentary: Cut sentences about process/intent
• Grand evaluations: Replace abstract praise with observable effects
• Over-justification: Allow judgment without explaining every reason
• Em/en dashes: Rewrite the sentence. Do not just swap for commas. Split into two sentences, fold the aside into the main clause, use a colon, reposition the aside, or drop it. Offer multiple options. Em dashes are strongly associated with LLM-generated text and should be eliminated.
• Self-posed rhetorical questions: "The result? Devastating." — rewrite as a declarative statement.
• "Serves as" dodge: Replace "serves as," "stands as," "marks," "represents" with "is" when it means the same thing.
• Dramatic countdown: "Not X. Not Y. Just Z." — state Z directly.
• False ranges: "From X to Y" where X and Y aren't on a real spectrum — name the items directly.
• Invented concept labels: Coined compound terms ("the supervision paradox") treated as established — describe the phenomenon in plain language instead.
• Superficial analyses: Tacked-on "-ing" phrases that inject shallow depth ("contributing to the region's rich cultural heritage," "highlighting its importance") — cut or replace with a specific claim.
• Semicolons: Check whether each semicolon could be a period. Two shorter sentences are usually clearer. Published sociology uses semicolons in ~10% of sentences, which is high.
• Transition word overuse: If "However," appears more than twice per page, suggest alternatives ("But," "Yet," or restructure). Total transition-led sentences should not exceed ~7%.
• Absolutist claims: Flag "all," "every," "none," "always" unless literally true. Suggest "most," "often," "typically," "in many cases."
• Prosecutorial tone: Flag "guilty," "verdict," "condemned," "exposed," "damning." Suggest "conclusion," "evidence indicates," "difficult to justify."
• Intensity check: For strong sentences, ask: Is intensity from evidence or wording? If from wording, revise down one notch.

Present sentence-level changes. Wait for user approval.

Level 4: Word

Final polish:

• Throat-clearing: "It is important to..." → [delete]
• Signposts: "Importantly," "Overall," → [delete]
• Evaluative adverbs: "convincingly demonstrates" → "demonstrates"
• Empty intensifiers: "reasonably comprehensive" → "comprehensive"
• Ability phrases: "managed to maintain" → "kept"
• Fancy-talk: "predicated upon" → "depends on"
• LLM vocabulary: Flag "delve," "tapestry," "landscape" (for domains), "leverage" (as verb), "robust," "streamline," "harness," "utilize," "synergy," "ecosystem" (outside biology), "navigate" (abstract), "paradigm" (outside Kuhn). Replace with plain equivalents.
• Magic adverbs: Flag "quietly," "deeply," "fundamentally," "remarkably," "arguably" when used to inject unearned significance. Cut or replace with evidence.
• Excessive praise: "thoughtful, rigorous, and sophisticated" → "careful"

Present word-level changes. Wait for user approval.

Output

After all levels approved, the edited file IS the output — it was edited in place. Commit the final state: git add [file] && git commit -m "writing-editor: Level 4 (Word) complete".

For pasted text (no file), write the final edited text to edited-[timestamp].md.

Include a brief summary of changes at each level in the conversation.

Calibration

Goal: Prose that sounds specific, slightly uneven, and willing to assert judgments without narrating its own cleverness.

Not the goal: Perfect prose. Functional prose is human. Allow mild awkwardness.

Genre awareness: Detect the genre from the input and respect its conventions. Academic papers keep citations and hedging where warranted. Memos stay short. Blog posts can be conversational. Reports keep structure tight. Don't flatten genre differences — adapt the rules to the context.

Final test: Read aloud. If it sounds like a report when it should be an essay, or a template when it should be a memo — keep editing.