Process Web Content Before It Hits Your Vault, Not After

There are four ways to add AI to an Obsidian workflow. The trade-offs:

• Smart Connections plugin (most popular). Embeds your entire vault to enable AI search. Works well under 1,500 notes. At 3,000+ notes, initial embedding takes hours and any vault rename re-embeds. Uses your OpenAI API key. Failure mode: the plugin gets blamed for "Obsidian feels slow" on large vaults.
• Text Generator plugin. Templater-style AI generation inside notes via slash commands. Powerful but requires you to author and maintain prompt templates per use case. Failure mode: templates drift; you forget which template does what.
• External flows (Make.com, n8n, Zapier). Webhook from clipper → AI processor → Obsidian. Cloud-dependent, multiple paid services, debugging is painful when something breaks. Failure mode: when the flow breaks (it will), you don't notice for weeks.
• Pre-clip processing with Prompt Anything Pro. Process the article in the browser. Decide what to clip. Clip only the structured output. Failure mode: requires the discipline to process before clipping — but the upside is your vault stays high-signal forever.

The case for the fourth approach: 80% of clipped articles are never re-read. Pre-clip processing turns "clipping" from a discovery act into a curation act. Only the articles that survive your AI-augmented pre-clip review hit the vault. The vault stays a high-signal asset; the noise stays in the browser tab where you closed it.

This guide focuses on the fourth approach. The Templater integration (step 6) covers the post-clip refinement path for notes that need additional processing later — covers cases where Smart Connections plugin would otherwise be your only option.

The 2026 clipper landscape:

• Obsidian Web Clipper (official, July 2025+). Browser extension by the Obsidian team. Clips as Markdown to a designated vault folder. Templates support YAML frontmatter. Use this — it's the most reliable.
• MarkDownload. Long-standing community extension. Conversion quality is solid. Use if you prefer per-clip naming control.
• Obsidian Clipper (community). Older fork. Skip in favour of the official.

Configure the Obsidian Web Clipper:

1. In the extension: set vault to your active vault.
2. Set default folder to 📥 Inbox (create this folder first in your vault).
3. Set template with YAML frontmatter — minimum: ---\nsource: {{url}}\nclipped: {{date}}\ntags: [inbox, unprocessed]\n---
4. The tags: [inbox, unprocessed] is critical — your Dataview / Bases queries surface these for processing review.

Why the Inbox convention: it creates a natural processing pipeline. Inbox → process with Prompt Anything Pro → permanent location. Without an Inbox, every clipped note instantly pollutes your topic folders.

This is the workflow that earns the most time per minute. On any article worth considering:

1. Open the article in your browser. Don't scan — at this stage you don't know if it's worth your time.
2. Cmd/Ctrl+A → select all article body (or use the browser's reader mode to strip nav/sidebar first).
3. Trigger Prompt Anything Pro.
4. Run this prompt verbatim: "You are evaluating an article for inclusion in a Zettelkasten knowledge base. Output: (1) One-sentence thesis. (2) Main arguments — 3-5 bullets. (3) Evidence quality — 1-2 sentences on whether the claims are sourced, anecdotal, or speculative. (4) Atomic concepts worth extracting as standalone notes — list 5-8 noun-phrase titles. (5) Verdict: KEEP (worth clipping in full + processing further), EXTRACT (clip only the atomic concepts, discard the rest), or SKIP (no Zettelkasten value). No preamble — return only the structured output."
5. If the verdict is SKIP: close the tab. The Prompt Anything Pro processing took 30 seconds and saved you from a permanent vault pollution.
6. If the verdict is EXTRACT: create one new note per atomic concept directly in your vault. Don't clip the article itself. Each new note title is the noun phrase from the AI output, body starts blank — you fill it in during reading.
7. If the verdict is KEEP: use the Obsidian Web Clipper to clip the article to Inbox. The AI-generated thesis + arguments paste at the top of the clipped note under ## Pre-Clip Summary. Tag it processed immediately so it doesn't reappear in your unprocessed queue.

The discipline payoff: across 10 articles you process this way, you might clip 2-3 in full and extract 3-4 atomic notes from another 2-3 articles. The other 4-5 don't enter your vault at all. Total processing time: ~15 minutes for 10 articles, vs hours of after-the-fact processing if you'd clipped all 10 raw.

For articles that scored EXTRACT or KEEP, the second prompt extracts atomic notes the way Luhmann actually structured Zettelkasten. Run this on the article body:

"Extract atomic Zettelkasten notes from this text. Each atomic note is a single independently-true claim that can stand alone without the source article. Output as Markdown — one note per heading. For each atomic note: (1) H2 heading with the claim as a noun-phrase title (NOT a verb phrase like 'how X works'). (2) Body: one-paragraph (max 3 sentences) standalone exposition of the claim. (3) Connections: a bulleted list under '## Connections' of 2-4 wikilink-format related concepts other notes in a Zettelkasten might link to — format as [[Related Concept]]. Extract 5-8 atomic notes from the article. No preamble — return only the Markdown."

Why this prompt works:

• Noun-phrase titles ("Compounding Interest" not "How compounding interest works") follow Luhmann's discipline — notes link by concept, not by question. Concept names compose into sentences naturally; question phrases don't.
• 3-sentence max forces atomic discipline. Anything longer is multiple notes pretending to be one.
• Wikilink suggestions jumpstart your vault's connection graph. You'll verify and accept some, reject others — but the suggestions surface non-obvious links between concepts.
• "Standalone without the source article" filters out atomic notes that are really just summary fragments. A true atomic note is true in isolation.

Paste the output into Obsidian: one new note per H2 heading, each named by its heading. Link them back to a "Source: [[Article Title]]" line in each. Your vault's connection graph builds itself.

Without a visibility layer, the Inbox folder silently accumulates. Two options to surface what needs processing:

Option A — Obsidian Bases (built-in, 2025+). Create a new Base in your vault root:

1. Cmd+P → "Create new base"
2. Set source: folder 📥 Inbox
3. Filter: tags contains "unprocessed"
4. Display columns: file name, clipped date, source URL
5. Sort: clipped date ascending (oldest first — fights recency bias)
6. Pin the Base to your sidebar

Option B — Dataview plugin. Drop this into your daily note template:

```dataview
TABLE source, clipped
FROM "📥 Inbox"
WHERE contains(tags, "unprocessed")
SORT clipped ASC
```

Both surface the same information: every Inbox note still marked unprocessed, oldest first. The processing flow becomes: open daily note → see the inbox queue → process each in Prompt Anything Pro using the prompts from steps 3-4 → remove the unprocessed tag → the entry drops off the list.

Without this visibility layer, the Inbox grows monotonically and the entire workflow eventually collapses. With it, the queue stays under control because the cost of not processing is visible every time you open your daily note.

For notes that landed in Inbox before you knew to pre-process, Templater is the cleanest in-Obsidian path to AI-augment them. Install Templater plugin, then add to your Templates/ folder a file called process-clipped-article.md:

<%* 
const noteContent = await app.vault.read(tp.config.target_file);
const body = noteContent.split('---').slice(2).join('---').trim();
const clipboard = body;
await navigator.clipboard.writeText(clipboard);
new Notice('Article body copied — paste into Prompt Anything Pro now');
-%>

The template, when executed, copies the article body to your clipboard so you can paste it into Prompt Anything Pro's popup directly. The workflow:

1. Open an unprocessed Inbox note
2. Run the process-clipped-article Templater command (bind to a shortcut)
3. Trigger Prompt Anything Pro on any page (Notion, Google Docs, a blank tab — anywhere)
4. Paste the article body into the Prompt Anything Pro input
5. Run the Zettelkasten Pre-Clip Filter prompt from step 3
6. Copy the AI output back, paste into the Obsidian note under ## Post-Clip Processing
7. Remove the unprocessed tag

The flow is 20-30 seconds per note — slower than pre-clip processing (because you're processing a note you already clipped) but faster than reading every Inbox note in full.

Common mistakes that pollute vaults despite this workflow:

• Raw AI output without your edits. The AI generated atomic notes are starting points — they're not yours until you've verified them against the source, modified phrasing to match your mental model, and added your own connections. Notes you haven't personally touched are vault noise.
• "Interesting" articles with no atomic concepts. If the AI returned 0-1 atomic concepts in step 3, the article is editorial or opinion — clip the verdict to a daily journal, not a permanent note. Permanent notes need standalone-true claims.
• News articles. Almost never warrant permanent vault entries. Time-bound claims (Q3 2026 economic data) become stale within weeks. If genuinely important, extract one atomic note ("recession starting Q3 2026" is the claim) and link to the source URL — but don't clip the article.
• YouTube transcripts and podcast notes. Often massive, often padded. Run the EXTRACT verdict more aggressively — typical 1-hour podcast yields 0-2 atomic concepts worth keeping, not 30. If you find yourself clipping every podcast in full, your verdict prompt isn't calibrated to your knowledge needs.
• Search-result pages, Reddit threads, social media. Almost never atomic-concept material. The signal-to-noise ratio is too low. If genuinely useful, save the URL to a daily journal entry — not as a permanent note.

The vault's value is concentration. 3,000 high-signal atomic notes outperform 30,000 mixed-quality clipped articles. The pre-clip workflow exists to keep the ratio favourable.