How We Built a 100% Client-Side C2PA Metadata Scrubber, Invisible Unicode Cleaner & NotebookLM Bundler in the Browser

As AI workflows matured in 2026, three technical bottlenecks hit developers, creators, and enterprise teams simultaneously:

1. Binary C2PA & JUMBF Provenance Flags: Even minor edits using Photoshop Generative Fill or Lightroom AI Denoise embed a cryptographic urn:c2pa JUMBF box and IPTC trainedAlgorithmicMedia tag—causing LinkedIn, Instagram, and TikTok to slap an automatic "Made with AI" badge on brand assets.
2. Hidden Zero-Width Unicode Characters (U+200B, U+202F): Copying text from LLM chat interfaces injects narrow no-break spaces and zero-width joiners that break legacy ATS parsers (like Workday and Greenhouse).
3. The 50-Source Limit in Google NotebookLM & 20-File Cap in Claude: Even though a single NotebookLM source accepts up to 500,000 words, uploading 50 short 3-page PDFs or Markdown files exhausts all 50 slots immediately.

To solve all three without ever uploading private files or confidential prompts to a cloud server, we built AIScrubbers.com — a 100% client-side browser utility suite.

---

1. Parsing JPEG APP11 JUMBF & PNG tEXt/iTXt/caBX Chunks in Pure JS

In PNG files, Stable Diffusion, ComfyUI, and Midjourney store entire prompt graphs inside uncompressed tEXt and iTXt chunks, while C2PA manifests reside inside caBX or jumb containers.

Our C2PA & AI Provenance Inspector reads the file's ArrayBuffer locally in browser RAM, parses chunk headers, and flags any C2PA or generator signatures. Then, our C2PA Metadata Remover reconstructs a clean pixel buffer via HTML5 Canvas—and for print creators on Amazon KDP or Etsy, our 300 DPI Print Normalizer injects an IEEE 802.3 CRC32-verified pHYs chunk (11811 pixels/meter = 300 DPI).

---

2. Bypassing Google NotebookLM's 50-Source Cap with XML Context Bundling

Anthropic's prompt engineering guide recommends wrapping multi-document context in structured <documents> and <document index="n"><source>filename</source> tags.

Using our NotebookLM & Claude 100+ File Merger, you can drop 100+ PDFs, .docx files, CSVs, and code files into your browser and compile a single indexed Master Markdown or PDF file—turning 100 files into 1 source slot while keeping pinpoint source citations intact.

---

3. Reversible Client-Side PII Redaction & llms.txt Generation

• Prompt Privacy Scrubber (PII Redactor): Tokenizes emails, SSNs, phones, and API keys ([EMAIL_1], [API_KEY_1]) locally in browser RAM before you paste into ChatGPT, and restores the real values into the AI's reply in 1 click.
• ChatGPT Invisible Watermark Cleaner: Strips U+200B, U+202F, U+FEFF, and curly quotes.
• Official llms.txt Generator & AI Crawler robots.txt Configurator: Allow OAI-SearchBot and PerplexityBot for AI search citations while blocking training scrapers.

Try the full 16-tool suite free (zero signups, zero server uploads) at https://aiscrubbers.com or view our documentation at GitHub.