Anti Ai Editor

Instructions

Phase 1: ASSESS

Goal: Read file, identify skip zones, scan for AI patterns.

Step 1: Read and classify the file

Read the target file. Identify file type (blog post, docs, README). Skip frontmatter (YAML between --- markers), code blocks, inline code, and blockquotes -- edits to these zones corrupt structure and would corrupt structure.

Auto-detect the content profile (linkedin, blog, technical-blog, investor-email, docs, casual) per references/context-profiles.md. Apply profile-specific tolerance rules throughout the scan.

If a voice profile is specified, also check voice-specific anti-patterns alongside the standard categories.

Step 2: Scan for issues by category

Some flagged words are appropriate in technical contexts. "Leverage" in "Use a lever to leverage mechanical advantage" is correct -- only flag words when used as corporate-speak, not in their literal or technical sense.

Step 3: Count and classify issues

Record each issue with line number, category, and severity weight:

• AI Cliche (Tier 1): weight 3
• News AI Tell (Tier 1-News): weight 3 (pseudo-profound, philosophizing, meta-significance)
• Copula Avoidance (Tier 1b): weight 3
• Meta-commentary: weight 2
• Dangling -ing clause (Tier 2b): weight 2
• Significance puffery (Tier 2c): weight 2
• Generic positive conclusion (Tier 2d): weight 2
• Dramatic AI rhythm (Tier 1-News): weight 2
• Structural issue: weight 2
• Fluff phrase: weight 1
• Passive voice: weight 1
• Redundant modifier: weight 1
• Curly quotes (Tier 3b): weight 1
• Novelty inflation (Tier 1g): weight 3
• Synonym cycling (Tier 1h): weight 2
• False concession (Tier 2e): weight 2
• Emotional flatline (Tier 2f): weight 2
• Reasoning chain artifact (Tier 2g): weight 2
• Parenthetical hedging (Tier 3c): weight 1
• Dash-as-separator (style): weight 2

Gate: Issues documented with line numbers and categories. Total severity score calculated. Proceed only when gate passes.

Phase 2: DECIDE

Goal: Determine editing approach based on severity.

Step 1: Choose approach by issue count

Step 1b: Rewrite-vs-patch threshold -- If severity score exceeds 30 AND 3+ distinct pattern categories are flagged AND structural rhythm is uniform, advise the user to consider a full rewrite rather than patching individual issues. Patching high-density AI text often introduces new patterns while fixing old ones.

Step 2: Prioritize fixes

1. Structural Issues (affect overall readability)
2. AI Cliches (most obvious tells)
3. Meta-commentary (usually removable)
4. Passive Voice (case-by-case judgment)

Every fix must be the minimum change needed. Multiple small edits beat one big rewrite because rewrites lose author voice and may introduce new AI patterns. Every detected issue must include a specific replacement -- reporting "Contains AI-sounding language" without a concrete fix is useless.

Step 3: Wabi-sabi check

Before proposing any fix, ask: "Would removing this imperfection make it sound MORE robotic?" If yes, preserve it. Preserve:

• Run-on sentences that convey enthusiasm
• Fragment punches that create rhythm
• Loose punctuation that matches conversational flow
• Self-corrections mid-thought ("well, actually...")

Natural informal language like "So basically" in a casual blog post is spoken rhythm, not an AI pattern. Only remove patterns that are AI-generated, not patterns that are merely informal.

Gate: Approach selected. Fixes prioritized. Wabi-sabi exceptions noted. Proceed only when gate passes.

Phase 3: EDIT

Goal: Generate edit report, get confirmation, apply changes.

Step 1: Generate the edit report

Show before/after for every modification with the reason -- always show before/after for every modification with the reason.

=================================================================
 ANTI-AI EDIT: [filename]
=================================================================

 ISSUES FOUND: [total]
   AI Cliches: [count]
   Passive Voice: [count]
   Structural: [count]
   Meta-commentary: [count]

 CHANGES:

 Line [N]:
   - "[original text]"
   + "[replacement text]"
   Reason: [specific explanation]

 [Continue for all changes]

=================================================================
 PREVIEW
=================================================================
[Show complete edited content]

=================================================================
 Apply changes? [Waiting for confirmation]
=================================================================

Style edits must preserve what the content says. When fixing "This solution robustly handles edge cases", write "This solution handles edge cases reliably" -- fix the style word, keep the technical meaning intact. If removing a flagged word would lose meaningful information, rephrase rather than delete.

Step 2: Apply changes after confirmation

Use the Edit tool for each change. Verify each edit applied correctly.

Gate: All changes applied. File re-read to confirm no corruption. Proceed only when gate passes.

Phase 4: VERIFY

Goal: Confirm edits preserved meaning and improved naturalness.

Step 1: Re-read edited file completely

Step 2: Verify no meaning was lost or changed

Step 3: Verify no new AI patterns were introduced by edits

Step 4: Confirm frontmatter and code blocks are untouched

Step 5: Report final summary

## Edit Summary
File: [path]
Issues Found: [count]
Issues Fixed: [count]
Issues Skipped: [count with reasons]
Meaning Preserved: Yes/No

Gate: All verification steps pass. Edit is complete.

Reference Material

Examples

Example 1: Blog Post (Heavy Editing)

User says: "De-AI this blog post" Actions:

1. Read file, skip frontmatter, scan all categories (ASSESS)
2. Score 22 -- systematic paragraph-by-paragraph approach (DECIDE)
3. Generate report with 10 changes, show preview, apply after confirmation (EDIT)
4. Re-read, verify meaning preserved, no new AI patterns (VERIFY) Result: 67% shorter intro, all AI cliches removed, voice preserved

Example 2: Technical Docs (Light Editing)

User says: "Check this for AI patterns" Actions:

1. Read file, identify technical context, scan for patterns (ASSESS)
2. Score 7 -- targeted fixes only, preserve technical terms (DECIDE)
3. Replace "utilizes" with "uses", remove throat-clearing, show preview (EDIT)
4. Verify technical accuracy unchanged (VERIFY) Result: Clearer prose, same information, technical terms untouched

Error Handling

Error: "File Not Found"

Cause: Path incorrect or file does not exist Solution:

1. Verify path with ls -la [path]
2. Use glob pattern to search: Glob **/*.md
3. Confirm correct working directory

Error: "No Issues Found"

Cause: Content is already natural, or scanner missed patterns Solution:

1. Report "Content appears natural -- no AI patterns detected"
2. Show sentence length statistics for manual verification
3. Check structural patterns (monotony, list overuse) even if no word-level flags

Error: "Frontmatter Corrupted After Edit"

Cause: Edit tool matched content inside YAML frontmatter Solution:

1. Fall back to treating entire file as content
2. Re-read file to verify YAML integrity
3. If corrupted, restore from git: git checkout -- [file]

References

• ${CLAUDE_SKILL_DIR}/references/cliche-replacements.md: Complete list of 80+ AI phrases with replacements
• ${CLAUDE_SKILL_DIR}/references/detection-patterns.md: Regex patterns for automated detection
• ${CLAUDE_SKILL_DIR}/references/detection-rules.md: Inline detection rules and structural checks
• ${CLAUDE_SKILL_DIR}/references/context-profiles.md: Content type profiles and tolerance matrix
• ${CLAUDE_SKILL_DIR}/references/examples.md: Before/after examples from real edits