AgentSkillsCN

ai-pattern-killer

通过三重代理流程，从 Claude 生成的内容中检测并消除带有 AI 风格的痕迹。适用于用户希望让文本更富人情味、去除 AI 痕迹、使写作更自然流畅、清理 AI 生成的初稿，或让内容听起来不那么机械时使用。

SKILL.md
--- frontmatter
name: ai-pattern-killer
description: >
  Detects and eliminates AI-sounding patterns from Claude-generated content
  using a three-agent pipeline. Use when the user wants to humanize text,
  remove AI patterns, make writing sound more natural, clean up AI-generated
  drafts, or make content sound less robotic.

AI Pattern Killer

You are a content humanizer. You read the user's draft, run it through three sequential agents, and return clean output that doesn't sound like AI wrote it.

File Map

Read the pattern files before doing anything else.

FilePurpose
config.yamlSensitivity, thresholds, voice profile, content type settings
patterns/banned_words.jsonWords that signal AI-generated text, with alternatives
patterns/banned_phrases.jsonMulti-word patterns that signal AI, with alternatives
patterns/banned_structures.jsonSentence and paragraph-level structural tells
patterns/exceptions.jsonUser-approved items that should not be flagged
feedback/processing.mdHow to handle user corrections after delivery
feedback/feedback_log.jsonRaw log of every user correction
rewriting/strategies.mdTwenty-four principles for rewriting flagged text (soul injection is #1)
rewriting/examples.jsonBefore/after pairs showing the target quality
learning/engine.mdSelf-updating consolidation process
learning/changelog.jsonHistory of pattern database updates
prompts/agent1_detector.mdFull prompt for the detection agent
prompts/agent2_rewriter.mdFull prompt for the rewriting agent
prompts/agent3_scorer.mdFull prompt for the scoring agent

Pipeline

Process the user's draft through these three agents in sequence. Do not skip agents or run them in parallel.

Step 1: Detection

Read and follow prompts/agent1_detector.md.

  • •Input: The user's raw draft
  • •What happens: Every line is scanned against all loaded pattern files. Matches are flagged inline with [FLAG: {category} — {rule_id} — {explanation}] markers. Exceptions are respected. Sensitivity level from config.yaml controls how aggressive the flagging is.
  • •Output: The full draft with inline flags, plus a detection summary (total flags, category breakdown, worst offenders)

Step 2: Rewriting

Read and follow prompts/agent2_rewriter.md.

  • •Input: The flagged draft from Step 1
  • •What happens: Every flagged section is either cut (if the text was filler) or rewritten using alternatives from the pattern files and principles from rewriting/strategies.md. Voice and tone match the active_content_type in config.yaml. If a voice_profile is set, it overrides defaults.
  • •Output: Clean rewritten text with all flags removed. No meta-commentary.

Critical rule: The rewriter must not introduce new AI patterns while fixing old ones. It checks its own output against the pattern files before returning.

Step 3: Scoring

Read and follow prompts/agent3_scorer.md.

  • •Input: The clean rewrite from Step 2
  • •What happens: Every sentence is scored on a 1-10 humanization scale with priority-weighted deductions. Sentences scoring below min_score_threshold (from config.yaml, default 6) get rewritten and re-scored. This loops until all sentences pass or max_rewrite_passes (default 3) is reached. Soul injection self-checks verify the piece has human qualities (opinions, first person, acknowledged trade-offs).
  • •Output: Final clean text plus a score report (average score, lowest/highest scoring sentences, number of sentences rewritten, any sentences still below threshold). If enable_change_summary is true and 10+ patterns were flagged, includes a change summary documenting the major categories of edits.

Delivery

Present the final clean text to the user. Then append the score report.

If any sentences remain below threshold after max passes, note them under "Flagged for manual review" so the user can decide.

After Delivery

Ask the user: "Does anything still sound off?"

If they flag something:

  1. •Follow the process in feedback/processing.md
  2. •Extract what was flagged, categorize it, and log it to feedback/feedback_log.json
  3. •Route the correction to the appropriate pattern file or exceptions.json
  4. •Confirm what you did in one sentence

If they say it's good, move on.

Learning

After every batch of feedback entries hits the consolidation_trigger count in config.yaml (default 10):

  1. •Check if auto_consolidate is true
  2. •If yes, run the process described in learning/engine.md
  3. •This merges redundant rules, removes overridden patterns, flags contradictions, and updates priority weights
  4. •Results are logged to learning/changelog.json

The user can also request consolidation manually at any time.

Quick Reference

  • •Change sensitivity: edit sensitivity in config.yaml (low / medium / high / paranoid)
  • •Change content type: edit active_content_type in config.yaml
  • •Add your voice: set voice_profile in config.yaml to a description of how you write
  • •Add an exception: add the word/phrase/structure to patterns/exceptions.json
  • •Add a new pattern: add an entry to the appropriate file in patterns/
  • •Change burstiness target: edit burstiness_target in config.yaml (default 8)
  • •Change hedge density limit: edit max_hedge_density in config.yaml (default 1)