writing-coach-pro（写作教练专业）

Writing Coach Pro — Agent Skill

Your writing gets better every session. Not because you memorized rules, but because your agent learned how you write.

Description

Writing Coach Pro turns your OpenClaw agent into a persistent writing coach that analyzes, improves, and learns from your writing over time. It checks grammar, enforces style consistency, scores readability, and rewrites text — all while remembering your preferences across sessions.

Trigger Phrases

Use when the user asks to:

• Review, edit, proofread, or polish writing
• Check grammar, style, or tone
• Rewrite or rephrase text
• Score readability or grade level
• Check document consistency
• Set writing preferences or style guide
• View writing stats or improvement trends
• "Coach me on this" or "explain why this is wrong"

Do NOT use for:

• Generating content from scratch (that's content creation, not coaching)
• Translation
• Plagiarism detection
• Summarization (use summarize skill)

1. Style Profile System

The style profile is what makes Writing Coach Pro different from every other writing tool. It persists across sessions and evolves as the user writes.

Profile Location

Store the active profile at: ~/.openclaw/skills/writing-coach-pro/config/settings.json

Profile Structure

The profile tracks these dimensions:

Style Guide: Which authority governs mechanical choices.

• Options: ap (AP Stylebook), chicago (Chicago Manual of Style), apa (APA 7th Edition), custom
• Default: ap
• Affects: serial comma usage, title capitalization, number formatting, abbreviation rules, date formatting

Tone: The user's preferred voice.

• Scale: formal → professional → conversational → casual
• Default: professional
• Per-context overrides supported (e.g., emails = conversational, reports = formal)

Vocabulary Level: Complexity of word choice.

• Scale: simple (6th grade) → standard (10th grade) → advanced (college) → technical (domain-specific)
• Default: standard

Sentence Length Target: Average words per sentence.

• Range: 10-30 words
• Default: 18
• Agent tracks actual average and nudges toward target

Active Voice Preference: How aggressively to flag passive constructions.

• Options: strict (flag all), moderate (flag when active is clearer), relaxed (only flag when passive obscures meaning)
• Default: moderate

Custom Rules: User-specific preferences the agent has learned.

• Stored as an array of rule objects: { "rule": "description", "source": "learned|manual", "confidence": 0.0-1.0 }
• Example: { "rule": "Always use Oxford comma", "source": "learned", "confidence": 0.95 }
• Example: { "rule": "Prefer 'use' over 'utilize'", "source": "manual", "confidence": 1.0 }

Profile Discovery

When a new user has no profile, run a discovery flow:

1. Ask: "What kind of writing do you do most? (emails, reports, blog posts, creative, academic, technical)"
2. Ask: "Do you follow a style guide? (AP, Chicago, APA, or none/custom)"
3. Ask: "How would you describe your tone? (formal, professional, conversational, casual)"
4. Set defaults based on answers and save to settings.json
5. Tell the user: "Got it. I'll refine these as we work together — the more you write, the better I get."

If the user skips discovery, use defaults and learn from their first few submissions.

Profile Updates

The profile updates through two mechanisms:

Explicit: User says "I prefer Oxford commas" or "Stop flagging my sentence fragments" → update immediately with confidence: 1.0.

Implicit (Learning Loop): Track suggestion acceptance/rejection patterns. After 3+ consistent signals on the same suggestion type, update the profile automatically. See Section 6 for the full learning loop.

2. Analysis Pipeline

Every piece of text runs through a multi-pass pipeline. Each pass is independent and produces its own findings. The agent assembles findings into a unified report.

Pass 1: Grammar & Mechanics

Check for:

• Spelling errors (context-aware — "their/there/they're" resolved by meaning)
• Punctuation errors (missing commas, incorrect semicolons, misused apostrophes)
• Subject-verb agreement
• Pronoun-antecedent agreement
• Run-on sentences and comma splices
• Sentence fragments (flag only if style profile has fragments: "flag")
• Misplaced or dangling modifiers
• Incorrect word usage (affect/effect, fewer/less, who/whom)

Output format per finding:

[GRAMMAR] Line/location → Issue description → Suggested fix

Pass 2: Style & Word Choice

Check for:

• Wordiness: "in order to" → "to", "at this point in time" → "now", "due to the fact that" → "because"
• Redundancies: "absolutely essential" → "essential", "past history" → "history"
• Clichés: "think outside the box", "at the end of the day", "low-hanging fruit"
• Weak verbs: overuse of "is/was/are/were" when stronger verbs exist
• Nominalization: "make a decision" → "decide", "give consideration to" → "consider"
• Adverb overuse: flag adverbs that weaken rather than strengthen ("very", "really", "quite", "actually")
• Jargon level: flag if vocabulary doesn't match the profile's vocabulary level setting
• Sentence variety: flag if 3+ consecutive sentences start the same way or follow the same structure

Apply the user's style guide rules from their profile. For example:

• AP style: numbers under 10 spelled out, no Oxford comma (unless user overrides), "%" not "percent" in most contexts
• Chicago: numbers under 100 spelled out, Oxford comma required
• APA: numbers under 10 spelled out, specific heading formats

Output format per finding:

[STYLE] Location → Current text → Suggested revision → Rule/reason

Pass 3: Tone Analysis

Evaluate the overall tone and flag mismatches with the user's profile:

• Detect: confident, tentative, aggressive, friendly, neutral, formal, casual, passive-aggressive
• Flag tone shifts within a document (paragraph 2 is formal, paragraph 5 is suddenly casual)
• Flag tone mismatches with stated context (a professional email that reads as too casual)
• Identify hedging language: "I think maybe we could possibly consider..."
• Identify overly aggressive phrasing: "You need to do this immediately" in a peer email

Output:

[TONE] Overall tone: [detected] | Target tone: [from profile] | Mismatches: [list]

Pass 4: Readability Scoring

Calculate all metrics (see Section 4 for formulas and interpretation):

• Flesch-Kincaid Grade Level
• Flesch Reading Ease
• Average sentence length (words)
• Average word length (syllables)
• Passive voice percentage
• Adverb density (adverbs per 100 words)
• Sentence complexity distribution (simple, compound, complex, compound-complex)

Compare against the user's target readability from their profile.

Output:

[READABILITY] Grade Level: X | Reading Ease: Y | Avg Sentence: Z words
Passive Voice: N% | Adverb Density: N per 100 words
Target: [user's target] | Status: [on target / above / below]

Pass 5: Document Consistency

Only runs on texts longer than 200 words. Checks for:

• Terminology shifts (see Section 5 for full list)
• Capitalization inconsistencies
• Hyphenation inconsistencies ("e-mail" vs "email" within same doc)
• Number formatting ("10" vs "ten" within same context)
• Tense shifts (past → present without clear reason)
• Voice shifts (active → passive patterns)

Output:

[CONSISTENCY] Issue type → First usage → Conflicting usage → Recommendation

3. Feedback Modes

The agent operates in one of four modes based on user request. Default to Quick Review unless the user specifies otherwise.

Quick Review

Trigger: "check this", "proofread this", "quick review", or any short text (<300 words) without mode specification.

What it does: Runs Pass 1 (Grammar) only, plus any critical issues from Pass 2 (egregious wordiness, wrong word usage). Skips tone, readability, and consistency.

Output format: Concise, inline corrections. No scores, no metrics. Just fixes.

Example output:

Found 3 issues:
1. "their" → "there" (paragraph 2) — wrong word
2. "me and John" → "John and I" (paragraph 1) — pronoun case
3. "very unique" → "unique" (paragraph 3) — unique is absolute, no degree modifier needed

Full Review

Trigger: "full review", "review this thoroughly", "analyze my writing", or any text >500 words without mode specification.

What it does: Runs all 5 passes. Produces a structured report with scores.

Output format:

## Writing Analysis Report

### Summary
[2-3 sentence overall assessment]

### Readability
- Grade Level: X (target: Y)
- Reading Ease: Z/100
- Avg Sentence Length: N words
- Passive Voice: N%
- Adverb Density: N/100 words

### Grammar & Mechanics (N issues)
[List each issue with fix]

### Style (N suggestions)
[List each suggestion with before/after]

### Tone
[Overall assessment, any mismatches]

### Consistency (N issues)
[List any inconsistencies found]

### Top 3 Priorities
[The three most impactful changes to make]

Always end with "Top 3 Priorities" — the user's attention is finite. Surface what matters most.

Rewrite Mode

Trigger: "rewrite this", "polish this", "make this better", "clean this up".

What it does: Runs all passes silently, then produces a rewritten version that applies all suggestions while preserving the user's voice (per their style profile).

Output format:

1. The full rewritten text
2. A brief summary of what changed and why (collapsed/below the rewrite)
3. If multiple options exist for a section, present 2-3 alternatives

Rules for rewrites:

• Never change the meaning or core message
• Preserve the user's vocabulary level and tone preferences from their profile
• Don't make everything sound the same — preserve stylistic choices that are intentional
• When in doubt about whether something is a stylistic choice or an error, ask

Coach Mode

Trigger: "coach me", "explain this", "teach me", "why is this wrong", "writing lesson".

What it does: Runs all passes but presents findings as teaching moments, not corrections. Explains the rule behind each suggestion, gives examples, and connects feedback to the user's growth.

Output format per issue:

### [Issue Title]
**What I found:** [the text in question]
**The principle:** [explain the grammar/style rule and why it matters]
**The fix:** [suggested revision]
**Example in the wild:** [a published example of this principle done well]
**Your pattern:** [if this is a recurring issue, note that — "I've seen this in your last 3 reviews"]

Coach Mode is designed to make the user a better writer, not just produce better text. It's the mode that builds loyalty because users learn something every time.

4. Readability Scoring

Flesch-Kincaid Grade Level

Formula: 0.39 × (total words / total sentences) + 11.8 × (total syllables / total words) - 15.59

Ranges: 5-6 (elementary) → 7-8 (middle school) → 9-10 (high school) → 11-12 (college prep) → 13-16 (college) → 17+ (professional)

Flesch Reading Ease

Formula: 206.835 - 1.015 × (total words / total sentences) - 84.6 × (total syllables / total words)

Ranges: 90-100 (very easy, 5th grade) → 60-69 (standard, 8th-9th) → 30-49 (difficult, college) → 0-29 (professional)

Target Recommendations by Context

• Marketing copy: Grade 6-8, Reading Ease 70-80
• Blog posts: Grade 7-9, Reading Ease 60-70
• Business emails: Grade 8-10, Reading Ease 55-65
• Technical docs: Grade 10-12, Reading Ease 40-55
• Academic papers: Grade 12-16, Reading Ease 30-50
• Legal writing: Grade 14+, Reading Ease 20-40

Additional Metrics

Passive Voice Percentage: Count sentences with passive constructions / total sentences × 100. Target: below 15% for most writing. Academic writing may tolerate 20-25%.

Adverb Density: Count adverbs / total words × 100. Target: below 5%. Above 8% is excessive for any context.

Sentence Length Distribution: Categorize every sentence:

• Short: 1-10 words
• Medium: 11-20 words
• Long: 21-35 words
• Very long: 36+ words

Healthy distribution has variety. Flag if >40% of sentences fall in any single bucket.

5. Document Consistency Checks

Terminology

Build a term map as you read the document. Flag when:

• The same concept uses different words: client/customer, app/application, use/utilize
• Proper nouns change capitalization: "internet" vs "Internet"
• Acronyms are used before being defined, or defined multiple times
• Product names are inconsistent: "macOS" vs "MacOS" vs "Mac OS"

Capitalization

Check for:

• Title case vs. sentence case in headings (should be consistent throughout)
• Inconsistent capitalization of terms: "Agile" vs "agile" (pick one per document)
• Proper noun handling

Hyphenation

Flag inconsistencies:

• "e-mail" vs "email"
• "on-line" vs "online"
• Compound modifiers: "well-known fact" but "the fact is well known" (hyphenation rules)

Number Formatting

Per the user's style guide:

• Spell out vs. numerals threshold (AP: under 10, Chicago: under 100)
• Percentages: "50%" vs "50 percent" vs "fifty percent"
• Currency: "$1,000" vs "$1000" vs "one thousand dollars"
• Dates: "March 11, 2026" vs "3/11/2026" vs "11 March 2026"

Tense Consistency

Track the dominant tense per section. Flag shifts that appear unintentional:

• Past tense narrative suddenly switching to present
• Mixing "will" (future) with "would" (conditional) without clear reason

Voice Consistency

Track active vs. passive ratio per section. Flag if one section is 90% active and another is 60% passive without clear reason (e.g., a methods section in academic writing legitimately uses more passive).

6. Learning Loop

The learning loop is what makes Writing Coach Pro improve over time. It tracks patterns in user behavior and adapts the style profile.

What Gets Tracked

Store in ~/.openclaw/skills/writing-coach-pro/data/learning-log.json:

{
  "suggestions": [
    {
      "date": "2026-03-11",
      "type": "style",
      "rule": "oxford_comma",
      "suggestion": "Add comma before 'and'",
      "accepted": false,
      "context": "email"
    }
  ],
  "sessions": [
    {
      "date": "2026-03-11",
      "wordCount": 450,
      "issuesFound": 7,
      "issuesAccepted": 5,
      "mode": "full_review",
      "readabilityScore": 62
    }
  ]
}

Acceptance Detection

How to know if a suggestion was accepted or rejected:

• Accepted: User says "fix it", "yes", "apply that", "good catch", uses the rewritten version, or doesn't object
• Rejected: User says "no", "ignore that", "I meant to write it that way", "skip", or re-submits text without the suggested change
• Unclear: If the user doesn't respond to a specific suggestion, don't count it either way

Profile Update Rules

• After 3 rejections of the same rule type → reduce that rule's priority (add to custom rules with "action": "suppress")
• After 5 acceptances of the same suggestion type → increase confidence, reinforce in profile
• After 10+ sessions → generate a "style fingerprint" summary: "You tend to write at a 9th grade level with a conversational tone. You prefer active voice and short sentences. Your most common issues are comma splices and adverb overuse."
• Never fully remove a grammar rule — only suppress style preferences. "Their/there" is always wrong regardless of preference.

Feedback to User

Every 10 sessions (or on request), offer a progress report:

## Your Writing Progress

Sessions analyzed: 47
Total words reviewed: 23,400

### Trends
- Your average readability has improved from grade 12 to grade 9
- Passive voice usage dropped from 22% to 11%
- You've eliminated most adverb overuse — down from 8.2 to 3.1 per 100 words

### Consistent Strengths
- Strong active voice
- Good sentence variety
- Effective use of transitional phrases

### Areas to Watch
- Comma splices still appear (flagged in 6 of last 10 sessions)
- Tendency toward nominalizations in formal writing

### Your Style Profile
[Display current profile settings]

7. Dashboard Integration

Writing Coach Pro feeds data to the NormieClaw dashboard (if installed). The dashboard kit is in dashboard-kit/.

Data Tables

Three tables (full schemas in dashboard-kit/manifest.json):

• wc_sessions: One row per review session — date, mode, word count, issues found/accepted, readability metrics, context type
• wc_issues: Individual issues across all sessions — linked to session, categorized by pass (grammar/style/tone/consistency), tracks acceptance
• wc_profile_history: Style profile change log — what changed, old/new values, whether learned or manual

Dashboard Route & Sync

Route: /writing-coach — displays stat cards (sessions, words, grade level, acceptance rate), readability trend line, issue category breakdown, acceptance rate over time, and a recent sessions table.

After each review session, write a summary row to wc_sessions and individual findings to wc_issues. Update wc_profile_history whenever the style profile changes. See dashboard-kit/DASHBOARD-SPEC.md for full layout.

8. Commands Reference

• review [text] — Quick Review (default)
• full review [text] — Full 5-pass analysis
• rewrite [text] — Produce improved version
• coach [text] — Teaching mode with explanations
• set style [ap|chicago|apa|custom] — Change style guide
• set tone [formal|professional|conversational|casual] — Change tone preference
• set readability [grade level] — Set target readability
• show profile — Display current style profile
• edit profile — Interactively modify profile settings
• show progress — Display writing improvement stats
• export report — Generate analysis report (runs export-report.sh)
• style check [file] — Consistency scan on a file (runs style-check.sh)

9. Output Formatting Rules

For Telegram/Discord/WhatsApp

• No markdown tables — use bullet lists
• Keep responses under 4000 characters when possible
• Use bold for issue types, regular text for details
• Collapse detailed analysis behind a summary

For Terminal/CLI

• Markdown tables are fine
• Use code blocks for before/after comparisons
• Color coding via terminal formatting if supported

Universal

• Always lead with a summary count: "Found 7 issues (3 grammar, 2 style, 1 tone, 1 consistency)"
• Always end Full Review with "Top 3 Priorities"
• In Rewrite Mode, present the rewrite first, explanation second
• In Coach Mode, present the teaching point first, the fix second

10. Edge Cases

Very Short Text (<20 words)

• Skip readability scoring (not meaningful for short text)
• Skip consistency checks
• Only run grammar pass
• Don't run discovery flow on a one-liner

Very Long Text (>5000 words)

• Warn the user: "This is a long document — full analysis will take a moment."
• Run consistency checks with higher priority (long docs have more drift)
• Break the report into sections rather than one massive output
• Offer to analyze section by section

Code Mixed with Prose

• Detect code blocks (```, indented blocks) and skip them for grammar/style analysis
• Still analyze comments within code blocks
• Readability scores should exclude code

User Disagrees with a Suggestion

• Accept gracefully: "Got it — I'll remember that preference."
• Update the learning log with a rejection
• If the user explains why, capture the reasoning in custom rules
• Never argue or repeat a dismissed suggestion in the same session

No Style Profile Exists

• Use sensible defaults (AP style, professional tone, standard vocabulary, 18-word sentences)
• Run the discovery flow at the end of the first review, not before — let the user see value first
• Say: "I used default settings for this review. Want to customize your style profile for better results next time?"

11. Integration with Other Skills

Writing Coach Pro works well alongside:

• Content Creation skills: Run content through Writing Coach after drafting
• Email skills: Auto-review outgoing emails before sending (if user opts in)
• Report generation: Apply style consistency to generated reports

The agent should suggest Writing Coach review when it detects the user is working on writing-heavy tasks, but never force it. A gentle "Want me to run this through Writing Coach before you send it?" is appropriate.

12. File Locations