██████╗░░█████╗░███╗░░░███╗░░██████╗░░█████╗░██████╗░██████╗░██╗░░░██╗
██╔══██╗██╔══██╗████╗░████║░░██╔══██╗██╔══██╗██╔══██╗██╔══██╗╚██╗░██╔╝
██║░░██║██║░░██║██╔████╔██║░░██║░░██║███████║██║░░██║██║░░██║░╚████╔╝░
██║░░██║██║░░██║██║╚██╔╝██║░░██║░░██║██╔══██║██║░░██║██║░░██║░░╚██╔╝░░
██████╔╝╚█████╔╝██║░╚═╝░██║░░██████╔╝██║░░██║██████╔╝██████╔╝░░░██║░░░
╚═════╝░░╚════╝░╚═╝░░░░░╚═╝░░╚═════╝░╚═╝░░╚═╝╚═════╝░╚═════╝░░░░╚═╝░░░

E X T R A C T   A N Y T H I N G

A Manifest V3 Chrome extension that extracts structured data from sites that fight scraping. Currently:

• ChatGPT, Claude, Gemini, Google AI Mode, AI Studio, Perplexity chats → Markdown / Text / JSON / CSV
• LinkedIn experience pages (/in/{you}/details/experience/) → Markdown / Text / JSON / CSV (one row per role)
• Anything else → RawMode, a two-step generic extractor for any site not in the list above

Pure client-side. No server, no build step at runtime, no analytics. One vendored dependency: Defuddle (MIT, ~290 KB ESM bundle) powers RawMode's article parsing.

Project structure

extension/
├── manifest.json
├── src/
│   ├── background/                   # Thin service worker (lifecycle hooks only)
│   ├── content/                      # One extractor per supported host
│   │   ├── chatgpt.js
│   │   ├── claude.js
│   │   ├── gemini.js
│   │   ├── aimode.js                 # Google AI Mode (google.com/search?udm=50)
│   │   ├── aistudio.js
│   │   ├── perplexity.js
│   │   ├── rawmode.js                # Generic extractor for unsupported sites (on-demand)
│   │   └── linkedin.js               # /in/{slug}/details/experience/
│   ├── lib/
│   │   ├── schema.js                 # Conversation / Profile / Article types (kind discriminator)
│   │   ├── markdown.js               # HTML → Markdown converter (no deps)
│   │   └── defuddle.js               # Vendored Defuddle bundle, powers RawMode
│   ├── exporters/
│   │   └── exporters.js              # export{Markdown,Text,JSON,CSV} for conversations
│   │                                 # export{ProfileMarkdown,ProfileText,ProfileJSON,ProfileCSV} for profiles
│   │                                 # export{ArticleMarkdown,ArticleText,ArticleJSON,ArticleCSV} for RawMode
│   └── popup/
│       └── popup.html / .css / .js   # User-facing UI; branches on result kind
└── icons/

RawMode

On any site not in the registry the popup shows Unknown site. Using RawMode... and a single Analyze Page button. Clicking it injects a generic content script that runs a tiered extraction pipeline:

1. Defuddle — full-fidelity article parsing on a re-parsed DOM clone (Defuddle is destructive). Returns title, byline, siteName, language, publishedTime, excerpt, word count, plus cleaned HTML. Then lib/markdown.js converts the HTML to Markdown.
2. Semantic walker — first <main> / <article> / [role="main"] element, with nav/aside/header/footer/script stripped, run through htmlToMarkdown. Used when Defuddle returns no useful content.
3. Plain text — document.body.textContent as a last-ditch fallback. Always succeeds on a non-empty page.

The chosen tier is recorded in the export's extractorTier field so you can spot when a page fell back. The four format buttons (Markdown / Text / JSON / CSV) replace the Analyze button once analysis completes, and the result is cached in chrome.storage.session keyed by tab + URL. Re-opening the popup on an already-analyzed page skips the Analyze step: the cached result is shown immediately and the page is re-analyzed in the background, so content that changed since (e.g. a chat that kept going) is picked up. Navigating away or restarting the browser clears the cache.

Install (unpacked)

1. Open chrome://extensions.
2. Toggle Developer mode on (top-right).
3. Click Load unpacked and select the extension/ directory in this repo.
4. Pin the extension to the toolbar.
5. Open a supported page, click the icon, choose a format. The Download / Copy toggle above the format buttons decides whether the result is saved as a file or copied to the clipboard; the choice is remembered. The circular refresh button next to the status re-reads the page without closing the popup — use it when the conversation has moved on since the popup opened.

For LinkedIn specifically: open your (or another user's) profile, click Show all experience, and run DOM Daddy on the resulting …/details/experience/ page. The popup will offer a "Take me there" shortcut if you click it on the wrong sub-page.

Works on any Chromium browser (Chrome, Edge, Brave, Arc, etc.).

Architecture

Data flow

Popup opened
  -> popup.js sends { type: 'EXTRACT' } to the active tab
  -> matching content script returns { kind, ...data }
  -> popup branches on kind, runs the right exporter
  -> chrome.downloads saves the file, or the clipboard receives the text (Download / Copy toggle)

Refresh button -> re-runs the whole pass against the page as it is now

The shared schema (extension/src/lib/schema.js) is the contract between extractors and exporters. Three shapes today: Conversation (kind: 'conversation'), Profile (kind: 'profile') and Article (kind: 'article', RawMode).

Why dynamic import() in content scripts?

MV3 doesn't allow content scripts to be declared as ES modules. To still share markdown.js and schema.js across extractors, each content script does await import(chrome.runtime.getURL('src/lib/...js')). The shared files are listed under web_accessible_resources so they're loadable from the page context.

Known limitations

• Selectors drift. When a site reorganizes, only that site's content script needs to change. Stable anchors:
  • ChatGPT: [data-message-author-role] / [data-message-id] inside section[data-turn]. The thread is virtualized with per-turn placeholder slots (no [data-turn] until scrolled into view), so the extractor scrolls each slot into view top→bottom and harvests as turns mount. In-prose file chips / download buttons are kept as text.
  • Claude: [data-testid="transcript-row"] → [role="article"][aria-posinset]; user text in [data-testid="user-message"], assistant in [data-is-streaming]. The transcript is a windowed virtual list, so the extractor scrolls top→bottom and harvests rows by position until it has aria-setsize messages.
  • Gemini: user-query, model-response (Angular component tags). Gemini lazy-loads older turns when scrolled to the top, so the extractor scrolls up repeatedly until the turn count stops growing.
  • Google AI Mode: [data-scope-id="turn"] per turn; the query is the turn's <h2> ("You said: …"), the answer is [data-subtree="aimc"] [data-container-id="main-col"]. Headings are div[role=heading][aria-level] and are rewritten to <hN>; the disclaimer/share footer block is dropped. Only google.com/search URLs with udm=50 or an mtid count as AI Mode — every other Google page falls through to RawMode.
  • AI Studio: ms-chat-turn → .user-prompt-container / .model-prompt-container → .turn-content; reasoning in <ms-thought-chunk>. Uses CDK virtual scrolling, so the extractor scrolls the chat top→bottom and harvests each turn as it mounts.
  • Perplexity: [class~="group/user-bubble"] for user queries, [data-workflow-final-text] .prose for answers. The thread list is found as the smallest container holding every query and answer (not by fixed parent hops, which broke when Perplexity added wrapper divs). It is virtualized (off-screen turns are empty placeholders), so the extractor scrolls each slot into view and harvests as it mounts; citation pills are rewritten to links and the hidden sticky-header copy of each table is dropped.
  • LinkedIn: [componentkey^="entity-collection-item-"] per company entry; we parse innerText line-by-line and ignore hashed CSS classes entirely.
• Extraction scrolls the page. Every chat host virtualizes or lazy-loads its transcript, so the extractors scroll through the conversation themselves, wait for turns to mount, and restore your scroll position afterwards. A long thread can take 20–60 s; keep the tab in the foreground while it runs — a background tab renders no frames, so nothing mounts. If an export still comes up short, the browser console will show a [DOM Daddy] warning saying how many turns never mounted.
• Collapsed UIs lose data. "Show thinking" details and LinkedIn's …see more may collapse content. Expand before extracting. Claude's collapsed thinking is not in the DOM at all; only the "Thought for Ns" label is.
• LinkedIn React hydration race. The entity-collection-item-* entries on the experience page render after the load event fires. The popup polls for up to ~6 seconds while LinkedIn hydrates, so a fast re-open immediately after navigation will wait briefly rather than fail.
• LinkedIn +N skills. The "+N skills" overflow on roles can't be pulled without clicking the chip — we capture the visible skills and store the hidden count as hiddenSkillCount.
• Canvas / Artifacts (ChatGPT side panel, Claude artifacts) contents aren't captured. Generated-file chips, download links and artifact cards are recorded by name in the message's attachments and kept as text in the Markdown.
• No real chat timestamps. None of the chat hosts expose creation date or per-message timestamps in the DOM, so the date in the filename is the export date.

Filenames and download-manager extensions

Conversations: {source}-YYYYMMDD-{sessionId}.{ext}. Profiles: {source}-{slug}-YYYYMMDD.{ext}.

If the Save As dialog shows a different filename than what we suggested, another installed extension is hooking chrome.downloads.onDeterminingFilename and overwriting our suggestion (Chrome only honors the most recently installed listener — there's no override). The popup's green "Saved …" status surfaces the real on-disk filename, so you can compare.

Known offender: Suno Tracks Exporter. Disable it (or any other download-manager extension) while exporting if you need the suggested filename to land.

Extending

Agent-facing guidance (architecture, DOM anchors, how to validate against a saved page or a live tab) lives in AGENTS.md; CLAUDE.md is a symlink to it.

Add a new conversation host

1. Add the host to manifest.json (host_permissions, content_scripts, web_accessible_resources).
2. Add a SITES entry in extension/src/popup/popup.js with kind: 'conversation'.
3. Copy an existing chat extractor and rewrite its selector block.

Add a new structured-data host (profile-style)

1. Same manifest updates, plus a pageReady/pageHint if the data lives on a sub-page.
2. New extractor returning makeProfile(...) (or a new schema shape if the data isn't profile-like).
3. New exporters in exporters.js if a new kind needs different output formats.

Add a new format

Add an exportXxx(data, opts) function in exporters.js returning { filename, blob }, then wire a button in popup.html and a case in runExport() in popup.js.

License

Code is licensed under the Apache License 2.0 (see LICENSE) — you can use, modify, and redistribute it commercially, subject to the usual Apache obligations.

The name DOM Daddy, the mascot, and all files under extension/icons/ plus favicon.ico are © Trollefsen Labs and are not covered by the code license. Forks must rename and re-skin before publishing to the Chrome Web Store or any other distribution channel. Full carve-out in NOTICE.

Why this exists

A from-scratch alternative to closed-source export extensions, with the goal of (1) keeping all extracted data inside the browser and (2) being trivially auditable — no minified bundles, no server, no analytics.