detect-writing（检测书写）

detect-writing

Detect the current writing system using multi-signal extraction and output a canonical writing-system spec.

Context Files

• $JAAN_LEARN_DIR/jaan-to-detect-writing.learn.md - Past lessons (loaded in Pre-Execution)
• $JAAN_CONTEXT_DIR/tech.md - Tech stack (for framework-aware i18n scanning)
• $JAAN_TEMPLATES_DIR/jaan-to-detect-writing.template.md - Output template
• ${CLAUDE_PLUGIN_ROOT}/docs/extending/language-protocol.md - Language resolution protocol

Output path: $JAAN_OUTPUTS_DIR/detect/writing/ — flat files, overwritten each run (no IDs).

Input

Arguments: $ARGUMENTS — parsed in Step 0.0. Repository path and mode determined there.

Pre-Execution Protocol

MANDATORY — Read and execute ALL steps in: ${CLAUDE_PLUGIN_ROOT}/docs/extending/pre-execution-protocol.md Skill name: detect-writing Execute: Step 0 (Init Guard) → A (Load Lessons) → B (Resolve Template) → C (Offer Template Seeding)

Language Settings

Read and apply language protocol: ${CLAUDE_PLUGIN_ROOT}/docs/extending/language-protocol.md Override field for this skill: language_detect-writing

Standards Reference

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-shared-reference.md for Evidence Format (SARIF), Evidence ID Generation, Confidence Levels, Frontmatter Schema, Platform Detection, Document Structure, and Codebase Content Safety.

This skill's namespace: E-WRT-* (e.g., E-WRT-001, E-WRT-WEB-001) Tool name in frontmatter: detect-writing

PHASE 1: Detection (Read-Only)

Step 0.0: Parse Arguments

Arguments: $ARGUMENTS

Mode determination:

• If $ARGUMENTS contains --full as a standalone token → set run_depth = "full"
• Otherwise → set run_depth = "light"

Strip --full token from arguments. Set repo_path to remaining arguments (or current working directory if empty).

Thinking Mode

If run_depth == "full": ultrathink If run_depth == "light": megathink

Use extended reasoning for:

• NNg tone dimension scoring across string corpus
• UI copy classification and quality assessment
• i18n maturity level determination
• Terminology consistency analysis

Step 0: Detect Platforms

Purpose: Auto-detect platform structure and determine analysis scope (full vs partial).

Use Glob and Bash to identify platform folders:

Platform Patterns

(Same as detect-dev - see detect-dev Step 0 for full patterns table)

Detection Process

1. Check for monorepo markers: pnpm-workspace.yaml, lerna.json, nx.json, turbo.json
2. List top-level directories: ls -d */ | grep -Ev "node_modules|\.git|dist|build|\.next"
3. Match against platform patterns: Apply disambiguation rules
4. Handle detection results:

- No platforms → Single-platform: platforms = [{name: 'all', path: '.'}] - Platforms detected → Multi-platform: Ask user to select all or specific platforms

Writing System Applicability

For each platform, determine analysis scope:

Partial analysis includes:

• Error message detection and scoring (Step 4)
• Glossary extraction from error messages (Step 5 - partial)
• Localization detection for error messages (Step 6 - partial)
• Skips: UI copy classification (Step 2), full tone analysis (Step 3 reduced to error messages only)

UI Presence Check

# Check for UI component files
ui_files=$(find {platform.path} -type f \( -name "*.jsx" -o -name "*.tsx" -o -name "*.vue" -o -name "*.svelte" \) 2>/dev/null | head -n 1)

if [ -z "$ui_files" ]; then
  # No UI files - partial analysis mode
  analysis_mode = "partial"  # Error messages only
else
  analysis_mode = "full"     # Full writing system analysis
fi

Analysis Loop

For each platform in platforms:

1. Set current_platform = platform.name
2. Set base_path = platform.path
3. Determine analysis mode based on platform type and UI presence
4. Run detection steps per run_depth and analysis_mode:

- If run_depth == "full" AND analysis_mode == "full": Run Steps 1-7 - If run_depth == "full" AND analysis_mode == "partial": Run Steps 1, 3 (reduced), 4, 5, 6 (reduced). Skip Step 2, 7 unless content linting detected. - If run_depth == "light" AND analysis_mode == "full": Run Steps 1, 5 only (skip Steps 2, 3, 4, 6, 7) - If run_depth == "light" AND analysis_mode == "partial": Run Steps 1, 4, 5 only (skip Steps 2, 3, 6, 7)

1. Use platform-specific output paths in Step 9

Partial Analysis Output Notes:

• writing-system.md: Tone dimensions based on error messages only, with note about scope limitation
• ui-copy.md: Minimal "Not Applicable" file with informational finding
• samples.md: Minimal "Not Applicable" file or error message samples only

Note: If single-platform mode (platform.name == 'all'), output paths have NO suffix. If multi-platform mode, output paths include -{platform} suffix.

Step 1: String Inventory

i18n / Locale Files

Use framework-specific glob patterns:

Component Inline Text

• Grep for text in component props: label, title, message, description, placeholder, helperText, errorMessage
• Grep for JSX/TSX inline text between tags

If run_depth == "light" AND analysis_mode == "full": Skip Steps 2-4, 6-7. Proceed directly to Step 5 (i18n Maturity Assessment).

If run_depth == "light" AND analysis_mode == "partial": Skip Steps 2-3. Proceed directly to Step 4 (Error Message Quality Scoring), then Step 5, then skip Steps 6-7.

Step 2: UI Copy Classification

Classify discovered strings into 8 categories (Buttons/CTAs, Error messages, Empty states, Confirmation dialogs, Notifications/toasts, Onboarding, Form labels/helper, Loading states) using component-name glob patterns and variant/severity props.

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "UI Copy Classification" for category detection patterns and props to extract.

Step 3: NNg Tone Dimension Scoring

Score strings across 4 primary dimensions (Formality, Humor, Respectfulness, Enthusiasm) and 5 extended dimensions (Technical complexity, Verbosity, Directness, Empathy, Confidence), each on a 1-5 scale. Calculate consistency scores via standard deviation per dimension; flag outliers deviating >1.5 standard deviations.

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "NNg Tone Dimension Scoring" for dimension scales, detection signals, and consistency calculation.

Step 4: Error Message Quality Scoring

Apply 5-dimension weighted rubric (Clarity 25%, Specificity 20%, Actionability 25%, Tone 15%, Accessibility 15%) to each error message found. Flag messages matching automated heuristic thresholds (complexity, length, passive voice, blame language, visible error codes, missing action verbs).

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "Error Message Quality Scoring" for rubric details and heuristic flag criteria.

Step 5: i18n Maturity Assessment

Use the glob patterns from Step 1 to identify which i18n framework is in use. Assess ICU MessageFormat usage, RTL support, hardcoded string prevalence, string interpolation quality, and centralization. Rate maturity on a 0-5 scale (None through Excellence).

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "i18n Maturity Assessment" for ICU regex patterns, RTL detection methods, hardcoded string grep patterns, interpolation quality levels, centralization scoring rules, and the 0-5 maturity scale.

If run_depth == "light": Skip Steps 6-7. Proceed directly to Step 8 (Present Detection Summary).

Step 6: Terminology Extraction

Build a glossary using ISO-704-inspired methodology. Discover terms via TF-IDF and C-value analysis, detect semantic/syntactic/frequency-based inconsistencies, and output entries with preferred/admitted/deprecated statuses.

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "Terminology Extraction" for term discovery methods, inconsistency detection rules, and glossary entry YAML format.

Step 7: Content Governance Detection

Check for content governance signals: CODEOWNERS locale ownership, content linting tools, i18n keywords in PR templates, and CI translation checks.

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-writing-reference.md — "Content Governance Detection" for specific tools and patterns to check.

HARD STOP — Detection Summary & User Approval

Step 8: Present Detection Summary

If run_depth == "light":

WRITING SYSTEM DETECTION COMPLETE (Light Mode)
-------------------------------------------------

PLATFORM: {platform_name or 'all'}
ANALYSIS MODE: {Full/Partial (error messages only)}

STRING CORPUS: {n} strings analyzed across {n} files
LOCALES DETECTED: {list}

i18n MATURITY: Level {0-5} ({name})
{if analysis_mode == "partial":}
ERROR MESSAGE SCORE: {avg_score}/10

SEVERITY SUMMARY
  Critical: {n}  |  High: {n}  |  Medium: {n}  |  Low: {n}  |  Info: {n}

OVERALL SCORE: {score}/10

OUTPUT FILE (1):
  $JAAN_OUTPUTS_DIR/detect/writing/summary{-platform}.md

Note: Run with --full for NNg tone dimensions, UI copy classification,
glossary, and governance analysis (6 output files).

"Proceed with writing summary to $JAAN_OUTPUTS_DIR/detect/writing/? [y/n]"

If run_depth == "full":

WRITING SYSTEM DETECTION COMPLETE
-----------------------------------

PLATFORM: {platform_name or 'all'}
ANALYSIS MODE: {Full/Partial (error messages only)}

STRING CORPUS: {n} strings analyzed across {n} files
LOCALES DETECTED: {list}

TONE DIMENSIONS (NNg) {scope note if partial: "based on error messages only"}
  Formality:      {score}/5    Consistency: {stddev}
  Humor:          {score}/5    Consistency: {stddev}
  Respectfulness: {score}/5    Consistency: {stddev}
  Enthusiasm:     {score}/5    Consistency: {stddev}

UI COPY COVERAGE {show "N/A" if partial analysis}
  Buttons:      {n} strings    Error messages: {n} strings
  Empty states: {n} strings    Dialogs:        {n} strings
  Toasts:       {n} strings    Onboarding:     {n} strings
  Form labels:  {n} strings    Loading:        {n} strings

i18n MATURITY: Level {0-5} ({name})
ERROR MESSAGE SCORE: {avg_score}/10

SEVERITY SUMMARY
  Critical: {n}  |  High: {n}  |  Medium: {n}  |  Low: {n}  |  Info: {n}

OVERALL SCORE: {score}/10

OUTPUT FILES (6):
  $JAAN_OUTPUTS_DIR/detect/writing/writing-system{-platform}.md  - Voice + tone + consistency
  $JAAN_OUTPUTS_DIR/detect/writing/glossary{-platform}.md        - Terminology glossary
  $JAAN_OUTPUTS_DIR/detect/writing/ui-copy{-platform}.md         - UI copy classification {or "N/A" if partial}
  $JAAN_OUTPUTS_DIR/detect/writing/error-messages{-platform}.md  - Error message audit
  $JAAN_OUTPUTS_DIR/detect/writing/localization{-platform}.md    - i18n maturity assessment
  $JAAN_OUTPUTS_DIR/detect/writing/samples{-platform}.md         - Representative samples {or "N/A" if partial}

Note: {-platform} suffix only if multi-platform mode (e.g., -web, -backend). Single-platform mode has no suffix.
      Partial analysis mode (backend/cli) produces minimal "Not Applicable" files for ui-copy.md and samples.md.

"Proceed with writing 6 output files to $JAAN_OUTPUTS_DIR/detect/writing/? [y/n]"

Do NOT proceed to Phase 2 without explicit approval.

PHASE 2: Write Output Files

Step 9: Write to $JAAN_OUTPUTS_DIR/detect/writing/

Create directory $JAAN_OUTPUTS_DIR/detect/writing/ if it does not exist.

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/detect-shared-reference.md sections "Output Path Logic" and "Stale File Cleanup" for platform-specific suffix convention and run_depth cleanup rules.

If run_depth == "light": Write Single Summary File

Write one file: $JAAN_OUTPUTS_DIR/detect/writing/summary{suffix}.md

Contents:

1. Universal YAML frontmatter with platform field, findings_summary, and overall_score
2. Executive Summary — BLUF of writing system findings
3. String Corpus Overview — total strings, file count, categories found (from Step 1)
4. i18n Maturity — level (0-5), locales detected, evidence (from Step 5)
5. If analysis_mode == "partial": Error Message Quality Scores — rubric scoring results (from Step 4)
6. Top Findings — up to 5 highest-severity findings with evidence blocks
7. "Run with --full for NNg tone dimensions, UI copy classification, glossary, and governance analysis (6 output files)."

If run_depth == "full": Write 6 Output Files

Write 6 output files:

Note: {suffix} is empty for single-platform mode, or -{platform} for multi-platform mode.

Partial Analysis "Not Applicable" Files:

For platforms with analysis_mode == "partial" (backend/cli), create minimal files for ui-copy.md and samples.md:

---
findings_summary:
  informational: 1
overall_score: 10.0  # Nothing to assess
---

## Executive Summary

Platform '{platform}' does not contain UI components. Full writing system analysis is not applicable. This audit focuses on error messages only.

## Findings

### E-WRT-{PLATFORM}-001: UI Copy Analysis Not Applicable

**Severity**: Informational
**Confidence**: Confirmed (1.0)

**Description**: Platform type '{platform}' (backend/CLI) does not have UI copy. Writing system analysis is limited to error messages, validation strings, and API responses.

Each file MUST include:

1. Universal YAML frontmatter with platform field and findings_summary/overall_score
2. Executive Summary (with scope note if partial analysis)
3. Scope and Methodology (clearly state "Partial Analysis" if applicable)
4. Findings with evidence blocks (using E-WRT-{PLATFORM}-NNN or E-WRT-NNN IDs)
5. Recommendations

Step 9a: Seed Reconciliation

Reference: See ${CLAUDE_PLUGIN_ROOT}/docs/extending/seed-reconciliation-reference.md for comparison rules, discrepancy format, and auto-update protocol.

1. Read domain-relevant seed files: $JAAN_CONTEXT_DIR/tone-of-voice.template.md, $JAAN_CONTEXT_DIR/localization.template.md
2. Compare detection results against seed content (tone dimensions, voice characteristics, error message guidelines, i18n maturity, supported languages)
3. If discrepancies found:

- Display discrepancy table to user - Offer auto-updates for non-destructive changes: [y/n] - Suggest /jaan-to:learn-add commands for patterns worth documenting

1. If no discrepancies: display "Seed files are aligned with detection results."

Step 10: Capture Feedback

"Any feedback on the writing system detection? [y/n]"

If yes:

• Run /jaan-to:learn-add detect-writing "{feedback}"

Skill Alignment

• Two-phase workflow with HARD STOP for human approval
• Evidence-based findings with confidence scoring
• Fork-isolated execution (context: fork)
• Output to standardized $JAAN_OUTPUTS_DIR path

Definition of Done

If run_depth == "light":

• Single summary file written to $JAAN_OUTPUTS_DIR/detect/writing/summary{suffix}.md
• Universal YAML frontmatter with overall_score
• String corpus overview and i18n maturity included
• If partial analysis: error message quality scores included
• "--full" upsell note included
• User approved output

If run_depth == "full":

• All 6 output files written to $JAAN_OUTPUTS_DIR/detect/writing/
• Universal YAML frontmatter with platform field in every file
• Every finding has evidence block with correct ID format (E-WRT-NNN for single-platform, E-WRT-{PLATFORM}-NNN for multi-platform)
• NNg tone dimensions scored with consistency analysis (note scope if partial)
• UI copy classified into 8 categories (or "Not Applicable" if partial analysis)
• Error messages scored with weighted rubric (always included)
• i18n maturity rated 0-5 with evidence (scoped to error messages if partial)
• Glossary uses ISO-704 statuses (error terminology if partial)
• Output filenames match platform suffix convention (no suffix for single-platform, -{platform} suffix for multi-platform)
• If partial analysis mode (backend/cli), minimal "Not Applicable" files created for ui-copy.md and samples.md
• Confidence scores assigned to all findings
• User approved output
• Seed reconciliation check performed (discrepancies reported or alignment confirmed)