César — Give back to César

A Chrome extension that detects parasitic lead magnets on LinkedIn — posts that gate access to content the author didn't create.

The problem

Some LinkedIn "creators" take official documentation, open-source projects, or public tutorials, repackage them in a post, then hide the link behind engagement bait: "Comment 🔥 and I'll send you the resource!"

The link is public. The content isn't theirs. They're farming engagement on someone else's work.

César detects these posts and helps you share the original source directly.

How it works

César uses a 3-layer confidence scoring model:

Layer 1 — Regex pattern matching (max 50%)

Scans every post in your feed for two combined signals:

• Third-party attribution: the post references a product, tool, or project created by someone else
• Gating CTA: the post gates access behind comments, DMs, likes, or follows

Both axes must fire. A post that gates its own content (legitimate lead magnet) is not flagged.

Layer 2 — Behavioral signals (max 75%)

Four heuristics that don't require AI:

Signal	What it detects	Max points
Comment bait	60%+ of comments are single emojis/keywords (🔥, "interested", "me")	+15
Public URL	Post mentions a GitHub repo, npm package, or official docs site	+10
Author mismatch	Author mentions Google but doesn't work at Google	+10
Post structure	Hype intro → feature list → CTA hidden at end	+5

Layer 3 — LLM verification (max 99%)

If an API key is configured, posts above 30% are sent to an LLM that answers the real question: "Did the author create this content, or just repackage something public?"

The LLM also generates a suggested comment with the original source and the César signature.

Threshold logic

Score	Without API key	With API key
< 30%	Ignored	Ignored
30-49%	Nothing shown	LLM must confirm (pending badge)
50-75%	Badge shown (capped)	Badge shown + LLM enriches
75-99%	—	Only reachable with LLM confirmation

Supported LLM providers

Provider	Model	Cost (per 1M tokens)	Key prefix
Anthropic	Haiku 4.5	$1 / $5	sk-ant-
OpenAI	GPT-5 mini	$0.25 / $2	sk-
Google	Gemini 3 Flash	$0.50 / $3	AIza
xAI	Grok 4.1 Fast	$0.20 / $0.50	xai-

The provider is auto-detected from the API key prefix. Prompt caching is enabled for Anthropic (saves ~90% on repeated calls).

Installation

From source (developer mode)

git clone https://github.com/bastiensoret/Cesar.git
cd Cesar
npm install
npm run build

1. Open Chrome → chrome://extensions/
2. Enable Developer mode (toggle top right)
3. Click Load unpacked
4. Select the dist/ folder
5. Go to linkedin.com/feed and scroll

Configuration (optional)

Click the César icon in the Chrome toolbar:

1. Toggle Web search if you want AI-powered source lookup (requires API key)
2. Open Settings → select your provider → paste your API key
3. The provider is auto-detected from the key prefix

Without an API key, César works in regex-only mode (capped at 75%).

Architecture

Cesar/
├── manifest.json          # Chrome Extension Manifest V3
├── build.js               # esbuild bundler (dev/build/package)
├── src/
│   ├── content/           # Content script — detection engine + overlay injection
│   ├── background/        # Service worker — multi-LLM API routing + prompt caching
│   ├── popup/             # Extension popup UI (HTML + CSS + JS)
│   ├── debug/             # MAIN world bridge for console debugging
│   └── shared/            # Utilities shared across modules
├── static/
│   ├── icons/             # Extension icons (16, 48, 128)
│   └── overlay.css        # Dark blue badge styles
├── store-assets/          # Chrome Web Store assets (screenshots, promo tile)
├── tests/                 # Vitest unit tests with jsdom
└── dist/                  # Build output (gitignored)

Key design decisions

• Manifest V3: content scripts run in isolated world, debug.js runs in MAIN world for console access
• No backend: all API calls go directly from the extension to LLM providers. No server to maintain.
• Prompt caching: Anthropic's cache_control: ephemeral caches the system prompt across calls
• Two-axis detection: a post must have BOTH a gating CTA AND a third-party reference to be flagged
• Layered scoring: regex alone can never say "parasitic" — it can only say "suspicious"

Badge UI

When a post is flagged, César injects a dark blue badge with:

• Header: César name + reason (truncated 40 chars) + confidence % + AI/no-AI badge
• Detail rows: what's gated + original source (LLM mode) or detection signals (regex mode)
• Comment section (LLM mode, collapsible): pre-generated comment with source + "Detected by César 🏛️" signature
• Feedback buttons: Confirm / Partial / False positive

The badge appears at the bottom of the post — you read the post first, then see the verdict.

Debug

Open Chrome DevTools console (F12) on LinkedIn:

// Enable debug logging
__cesar.enableDebug()

// Force re-scan all posts
__cesar.rescan()

// Test regex detection on a text string
__cesar.test("Google just released something amazing! Comment 🔥 and I'll send you the link")

// View scan stats
__cesar.stats()

Debug logs show the full layer breakdown:

[César] L2 Score: 62% (cap 75% (no AI)) | Major company launch mentioned + Asks to comment an emoji
  L1 regex: TP=65 CTA=80 → 36%
  L2 behavioral: +26pts → comments +15, public URL +10, mismatch +10

Roadmap

Next up

• Web search integration (LLM with internet access for verified source URLs)

Community features

• Backend + database for storing detections
• Recidivist leaderboard (most flagged accounts)
• Community-submitted source directory
• Shared flagging (if 3+ users flag same post, instant detection)

Platform expansion

• Firefox support (WebExtension API)
• Landing page
• LinkedIn post template for spreading awareness

Contributing

This is an open-source community tool. Contributions welcome:

1. Fork the repo
2. Create a feature branch (git checkout -b feature/my-feature)
3. Test on your LinkedIn feed
4. Submit a PR with before/after screenshots

Areas where help is most needed

• Regex patterns: more languages (DE, ES, PT, NL), edge cases
• LinkedIn DOM selectors: these break frequently, need monitoring
• LLM prompt tuning: reducing false positives on legitimate lead magnets
• Comment pre-fill: LinkedIn's contenteditable is fragile, needs robust injection

License

MIT — Give back to César ⚡