Clement-2333/ai-clean-master

AI Clean Master (All-Carrier AI-Slop Cleaner & Humanizer)

1. Core Mission

AI-generated PPTs, infographics/posters, websites, and documents often reek of "machine pipeline plastic" (AI Slop), concentrated in two main ailments:

1. Visual & Typographic AI-Slop Micro-Copy and Cheap Decorative Frames (Visual & Micro-Copy Slop):
   • Top/Card Fake English & Cliché Kickers/Pill Badges: Like SECTION 01 · STRATEGIC OVERVIEW, CORE INSIGHT, ✨ Next-Gen AI Platform, SYSTEM ONLINE, or 核心赋能模块, usually wrapped in a rounded pill/badge or glowing border.
   • Redundant, Space-Filling Micro-Copy Inside Cards: The headline already says "Recognition speed improved 3x," but beneath it is a forced two-line 10pt / text-xs filler like "Through the introduction of efficient algorithmic mechanisms, a significant enhancement in recognition processing speed has been achieved."
   • Footer/Bottom Bar Grandiose Summaries & Disclaimer Fluff: Such as 本页核心洞察：构建全链路数字化闭环, Key Takeaway: Empowering seamless digital transformation, Note: Data for demonstration purposes only, AI watermarks, or fake command-line status bars.
   • Fake Decorative Micro-Data/Badges & Symbol Noise: Baseless +99.8%, INDEX-A, STEP 01 //, and rampant ✨🚀💡⚡ emojis with cheap purple/blue neon glows.
2. Linguistic Chinese-English AI Jargon & Cliché (Linguistic AI Buzzwords):
   • High-Frequency Chinese Jargon: "赋能" (empower), "抓手" (grip/pivot), "闭环" (closed loop), "全链路" (full-chain), "多维度" (multi-dimensional), "底层逻辑" (underlying logic), "顶层设计" (top-level design), "生态体系" (ecosystem), "范式转移" (paradigm shift), "协同共振" (synergistic resonance), "颗粒度" (granularity), "组合拳" (combination punch) and robotic four-character parallelism.
   • High-Frequency English Jargon: delve into, testament to, rich tapestry, seamless integration, paradigm shift, holistic ecosystem, empower, leverage, cutting-edge, game-changer, unlock the full potential.

Four Iron Rules of This Skill:

• Box-and-Text Removal: When removing decorative micro-copy, the pill/badge background box (Pill/Badge) and bottom bar (Footer Bar) must be removed along with the text – never leaving an empty colored block.
• Smart Visual Promotion: After deleting the redundant filler text inside a card, automatically enlarge the retained core viewpoint (e.g., PPT +2pt~+4pt, web text-base -> text-lg font-semibold), increase line spacing, or vertically center the text to prevent the layout from feeling empty.
• Adaptive Humanization: Consistently chop out Chinese and English buzzwords and jargon, and rewrite the content into natural, speakable language based on the target scenario (Pitch Competition, Engineering Defense, Classroom Teaching, Crisp Business).
• Non-Destructive Output: All modifications are output to a new *_cleaned.<ext> file – the original is never touched.

---

2. Cross-Platform Five-Step Workflow (Integrating grill-me Questioning & User Confirmation Gate)

Retain the absolute directory path containing this file as SKILL_DIR. When invoking scripts, expand ${SKILL_DIR} to the real absolute path.

Stage 1: Carrier Identification & AI-Slop Lesion Scanning (Scan & Diagnose)

Based on the target file or code type provided by the user, call the corresponding dedicated scanning engine (recognition criteria can be found in references/ai-slop-dictionary.md):

Carrier Type	Applicable Extensions	Scan Command
PPT Presentation	.pptx	python "${SKILL_DIR}/scripts/pptx_inspector.py" "<input.pptx>" --json "<scan.json>"
Website / Frontend Page	.html, .jsx, .tsx, .vue, .svg	python "${SKILL_DIR}/scripts/web_cleaner.py" --scan "<input_file>" --json "<scan.json>"
Image / Poster / Infographic	.png, .jpg, .jpeg, .webp	python "${SKILL_DIR}/scripts/image_cleaner.py" --scan "<input_img>" combined with multimodal vision to locate pixel coordinates bbox of all micro-copy, pill badges, and redundant side copy in the image
Document / Outline / Speech Script	.docx, .md, plain text	python "${SKILL_DIR}/scripts/doc_cleaner.py" --scan "<input_doc>" --json "<scan.json>"

[!NOTE]

• On Windows, if python3 triggers the WindowsApps stub (exit code 9009), uniformly use python.
• When in a purely cloud-based dialogue environment like Doubao Workspace without a local Python container, perform structured diagnosis and code/copy rewriting directly by referring to references/ai-slop-dictionary.md.

---

Stage 2: Grill-Me One-by-One Interactive Decision Interview

After scanning, modifying the file is strictly prohibited. You must enter the grill-me interview mode, following the decision tree to ask only one question at a time, and for each question, place the most recommended option first and label it (Recommended):

• Cross-Agent Tool Call Adaptation:
  • In hosts that support structured questioning tools (e.g., ask_question or AskUserQuestion) like Antigravity / Claude Code / Cursor / Trae / Qoder / WorkBuddy, prefer calling that tool to initiate a single-choice question to the user.
  • In environments without popup selection tools, like Codex CLI / Doubao Workspace, present only one question at a time in the dialogue box using a clear numbered list (1. (Recommended) ...), and wait for the user's replied number or opinion before asking the next question.

Resolve the following decision branches in order (if the user has already explicitly specified an item in their initial instruction, that question can be skipped):

1. Q1 — Target Scenario & "Human Tone" Gear (Tone & Scenario):
   • Automatically place the best-matching style based on the scan content first and label it (Recommended):
     • Pitch Competition / Startup Defense Style (Pitch & Competition): Highlight real pain points, "what we built ourselves," and measured results (reject empty talk of ecosystems and empowerment).
     • Engineering Training / Technical Engineering Style (Engineering & Technical): Verb-first + specific hardware/algorithm/parameters + quantified results, objective and concise.
     • Classroom Teaching / Class Meeting Speech Style (Teaching & Speech): Includes first/second-person subjects ("咱 们看" / "大家常遇到" / "Let's look at..."), short sentences, colloquial.
     • Crisp Business / Daily Report Style (Crisp Business): Conclusion first (one page, one human-readable sentence of the conclusion), using facts and specific actions instead of four-character adjectives.
2. Q2 — Micro-Copy & Decorative Color Block Cleanup Aggressiveness (Cleanup Aggressiveness):
   • (Recommended) Full Box-and-Text Cleanup + Font Promotion: Delete all top fake English/cliché micro-pills (along with their background boxes), bottom footer disclaimer fluff, and redundant filler micro-copy inside cards. Then enlarge and center the retained core viewpoint font size.
   • Moderate Cleanup: Only delete obvious fake English badges, footer disclaimers, and ✨🚀 emojis. Retain the subtitle structure inside cards but rewrite it into concise human language.
3. Q3 — Confirmation of Borderline Real Data Badges or Image Mode (Borderline Data / Image Mode):
   • If the scan finds micro-copy/badges with specific numbers (e.g., +38.5%, N=120), create a single question asking the user to confirm which are real business data (must be retained and promoted) and which are AI-generated decorative micro-data (must be box-and-text deleted).
   • If the target is an image (.png/.jpg), ask whether to adopt (Recommended) Pixel-level seamless micro-copy erasure on the original image + in-place redraw of human-large text, or Reconstruct into editable HTML/SVG/PPTX layers.

---

Stage 3: Output "Deletion + Humanization Rewrite + Visual Promotion" Comparison Table & Confirmation Gate

After completing the grill-me alignment, read references/layout-promotion-rules.md and references/ai-slop-dictionary.md, and present to the user an intuitive per-page/per-zone Before → After comparison preview table:

### 📋 AI Clean Master Cleanup & Humanization Preview

| Page/Zone | Action Type | Original Content (AI-Slop Lesion) | After Treatment (Humanized & Visually Promoted) |
|---|---|---|---|
| P1 Top | `🗑️ Box-Delete` | `[Pill Box #4] SECTION 01 · STRATEGIC OVERVIEW` | *Removed along with the light blue capsule background box* |
| P1 Card1 | `✂️ Delete micro-text + 🔼 Promote` | Headline: `多维赋能分选体系`<br>Micro-text(10pt): `通过全链路视觉感知赋能高效农业分选闭环` | Headline rewritten and promoted to **20pt centered**:<br>**`双目视觉 0.2 秒识别单果瑕疵`** (deleted the filler micro-text below) |
| P1 Bottom | `🗑️ Box-Delete` | `[Footer Bar #9] 本页核心洞察：实现高质量发展的抓手` | *Removed along with the bottom dark gray horizontal bar* |

• Confirmation Gate: After displaying the preview table, ask the user to confirm execution (or suggest adjustments). Only when the user explicitly confirms (or stated clearly in the initial request "Don't confirm, just give me a one-shot result") can you proceed to Stage 4.

---

Stage 4: Execute Box-and-Text Cleanup, Humanization Rewrite, and Smart Visual Promotion (Execute Cleanup & Promotion)

Based on the confirmed plan, call the carrier-specific cleanup engine to generate the *_cleaned.<ext> new file:

Route A: PPT Presentation (.pptx)

Write a JSON plan file and run:

python "${SKILL_DIR}/scripts/pptx_cleaner.py" "<input.pptx>" "<plan.json>" -o "<input_cleaned.pptx>"

Example plan.json structure:

{
  "slides": [
    {
      "slide_index": 1,
      "delete_shape_ids": [4, 5, 9, 10],
      "delete_paragraphs": [
        {"shape_id": 7, "paragraph_indices": [1, 2]}
      ],
      "rewrites": [
        {
          "shape_id": 7,
          "paragraph_index": 0,
          "new_text": "双目视觉 0.2 秒识别单果瑕疵",
          "promote_pt": 3.0,
          "vertical_center": true
        }
      ]
    }
  ]
}

Route B: Website & Frontend Pages (.html / .jsx / .tsx / .vue / .svg)

Write <web_plan.json> and run:

python "${SKILL_DIR}/scripts/web_cleaner.py" --apply "<web_plan.json>" "<input.html>" -o "<input_cleaned.html>"

Route C: Image / Poster / Infographic (.png / .jpg / .webp)

Write a zone erasure and redraw configuration <img_plan.json> (defining bbox: [x1, y1, x2, y2], action: "erase_only" | "erase_and_redraw", text, font_size for each zone), then run:

python "${SKILL_DIR}/scripts/image_cleaner.py" --apply "<img_plan.json>" "<input.png>" -o "<input_cleaned.png>"

Route D: Document & Rich Text (.docx / .md)

Write <doc_plan.json> and run:

python "${SKILL_DIR}/scripts/doc_cleaner.py" --apply "<doc_plan.json>" "<input.docx>" -o "<input_cleaned.docx>"

---

Stage 5: Delivery of Results & Cleanup Report (Delivery & Summary Report)

Verify that the generated *_cleaned.* file exists and has no format damage, then present a concise delivery report to the user:

1. New file path;
2. Cleanup battle statistics: Number of box-and-text deleted micro-copies/pill badges/footer bars, number of chopped Chinese and English AI jargon terms, and number of completed visual font promotions and humanization rewrites of core viewpoints.

---

3. Reference Documents & Script Index / References & Scripts

• references/ai-slop-dictionary.md: Chinese and English AI-slop micro-copy feature library, banned jargon list, and four-scenario humanization rewrite dictionary.
• references/layout-promotion-rules.md: Geometric judgment for box-and-text cleanup and font promotion specification for PPT/images/web pages.
• scripts/pptx_inspector.py & scripts/pptx_cleaner.py: .pptx scanning and cleanup/promotion engine.
• scripts/web_cleaner.py: HTML / React / Vue / SVG scanning and denoising/promotion engine.
• scripts/image_cleaner.py: Poster and infographic pixel-level seamless micro-copy erasure and large-text redraw engine.
• scripts/doc_cleaner.py: Word (.docx) and Markdown (.md) AI-slop removal engine.
• scripts/install_skill.py: One-click installation and adaptation script for mainstream cross-platform AI agents.