obsidian-web-clipper

name	obsidian-web-clipper
description	Author and debug Obsidian Web Clipper extension templates: template JSON, variables, filters, template logic, schema.org/CSS selectors, AI interpreter prompts, URL/schema triggers. Use when generating, importing, or fixing a clipper template, or matching one to a target site. Not for general scraping.
metadata	{"author":"nweii","version":"1.6.0"}

Obsidian Web Clipper Templates

Obsidian Web Clipper is a browser extension that saves web content to an Obsidian vault as Markdown notes. Templates define how pages are captured — what metadata to extract, how to format the note, and which sites to auto-match.

Templates are configured as JSON. Users can import/export individual templates or full settings backups. When generating templates, output valid JSON the user can import directly.

Single-page scope. A clip only sees the page it fired on — the DOM loaded in the active tab. The extension is not an agent; it can't crawl, follow links, or visit other URLs. Selector, schema, and interpreter variables all resolve against that one page. If a template needs data spread across multiple pages, it can't gather it in one clip — design around what's on the page being clipped.

Caution: full settings exports contain interpreter API keys in plaintext (interpreter_settings.providers[].apiKey). Never quote them; flag exports stored in version control.

Adapt to the user's vault conventions

Templates hold opinionated conventions for folder layout, frontmatter property names, category/tag conventions, and note body structure. Match the user's existing conventions rather than inventing.

Before drafting a template, understand:

• Vault-level docs (CLAUDE.md, AGENTS.md, READMEs)
• An existing templates folder, sample notes of the type being clipped, or a previously exported clipper settings file
• Any YAML property-sort configuration (Linter or similar)
• User preferences uncovered in discussion

When conventions aren't discoverable or if the user isn't sure what they want, tailor what you can and fall back to minimal vendor-neutral choices. Flag any assumptions you've made so the user can adjust before import.

If the vault has a property sort order, mirror that order in the template's properties array.

Template JSON Schema

{
  "schemaVersion": "0.1.0",
  "name": "Template Name",
  "behavior": "create",
  "noteNameFormat": "{{title}} - {{site}}",
  "noteContentFormat": "{{content}}",
  "vault": "VaultName",
  "path": "folder/subfolder",
  "context": "",
  "properties": [],
  "triggers": []
}

Field	Type	Description
schemaVersion	string	Always "0.1.0"
name	string	Display name shown in the extension
behavior	string	How the note is created (see below)
noteNameFormat	string	Filename for the note, supports variables
noteContentFormat	string	Note body content, supports variables
vault	string	Target vault name (optional — omit to use default)
path	string	Folder path within the vault (optional)
context	string	Limits what page content the interpreter AI sees (optional)
properties	array	Frontmatter properties (optional)
triggers	array	Auto-match rules for URLs/schemas (optional)

When exporting a single template, the JSON is the template object directly (no wrapper). When part of a full settings export, templates are stored as template_[id] keys.

Behaviors

Value	Description
create	Create a new note
append-daily	Append to today's daily note
prepend-daily	Prepend to today's daily note
append-specific	Append to a specific existing note
prepend-specific	Prepend to a specific existing note

Variables

All variables use {{variableName}} syntax. Filters chain with pipes: {{variable|filter1|filter2:"arg"}}. There are four kinds:

• Preset variables — built-in page metadata: {{title}}, {{url}}, {{author}}, {{site}}, {{published}}, {{description}}, {{content}} (full article markdown), {{date}}, {{highlights}}, etc. Look up the full list in the live docs.

• Schema variables — extract Schema.org JSON-LD from the page. Patterns:

  {{schema:name}}                    — first match anywhere
  {{schema:@Recipe:name}}            — scoped to a specific @type
  {{schema:author.name}}             — nested keys via dots
  {{schema:image[0].contentUrl}}     — array index
  {{schema:actors[*].name}}          — flatten arrays
  

  Schema-driven templates are usually the right choice for domain-specific clippers (recipes, films, books, jobs) since one schema:@Type trigger matches across many sites.

• Selector variables — pull content via CSS selectors when there's no schema:

  {{selector:h1}}                    — text content
  {{selector:img.hero?src}}          — attribute value
  {{selectorHtml:article}}           — raw HTML
  {{selectorHtml:body|markdown}}     — HTML → Markdown
  

• Interpreter variables — natural-language prompts evaluated by an LLM (requires the user has an LLM provider configured in extension settings):

  {{"a summary of the page"}}
  {{"3 tags describing this content"}}
  {{"return JSON array with fields: author, text"|map:item => item.text|join:"\n"}}
  

  The context field at the template top level controls what page content the AI sees. Use {{selectorHtml:#main}} or wrap structured context like <page>\nTitle: {{title}}\n{{content}}\n</page> to keep prompts focused.

Filters

Chain filters with |. Look up specific filter signatures in the live docs (see Reference at the bottom). The patterns and gotchas below are the operationally non-obvious parts.

Filter gotchas (verified via testing)

Minutiae that might trip up template authors:

• String concatenation with + doesn't work inside map callbacks. The expression parser treats + as an unexpected character and fails the template import with "Unexpected character '+' in template". Use template literals instead:

  ✗ |map:item => "- " + item               (parse error on import)
  ✓ |map:item => "- ${item}"               (template literal — correct)
  

• Built-in filters cannot be chained inside map. Flatten or transform inside map, then apply filters to the result outside:

  ✗ |map:item => item.text|trim            (filter not allowed inside map)
  ✓ |map:item => item.text|join:"\n"|trim  (filter chain outside map)
  

• For arrays of objects with multi-line per-item output, use map to flatten to flat keys, then template. Nested key access in template literals is not documented and unreliable; flattening first is the safe pattern:

  |map:item => ({name: item.author.name, body: item.reviewBody})
  |template:"${name}:\n> ${body}\n\n"
  

• duration outputs HH:mm:ss, not compact human format. For something like 45m or 1h 30m, skip the filter and chain replace on the raw ISO 8601 string:

  {{schema:@Recipe:prepTime|replace:"PT":""|replace:"H":"h "|replace:"M":"m"|trim}}
  → "PT1H30M" becomes "1h 30m"
  → "PT45M"   becomes "45m"
  

• |list works on string arrays directly — no need to map a "- " + item prefix yourself. Use |list for bullets, |list:numbered for numbered, |list:task for checkboxes.

Common Filter Chains

{{url|split:"?"|slice:0,1}}                              — strip query params
{{schema:actors[*].name|wikilink|slice:0,4|join}}         — first 4 actors as wikilinks
{{highlights|map:item => item.text|join:"\n\n"|blockquote}} — highlights as blockquote
{{date|date:"YYYY-MM-DD-ddd"}}                            — formatted date with day name
{{"return JSON..."|map:item => item.title|join:"\n"}}     — AI → structured → formatted

Template logic

Templates support Twig/Liquid-style logic in noteContentFormat, noteNameFormat, and property values: {% if %}/{% elseif %}/{% else %} conditionals, {% for %} loops (with a loop object: loop.index, loop.first, loop.last, etc.), {% set %} variable assignment, and ?? fallbacks ({{title ?? "Untitled"}}). Full syntax lives in Logic.md in the live docs (see Reference). The non-obvious parts:

• Evaluation order: template logic runs first, interpreter prompts after. Logic can construct a prompt dynamically, but a prompt's result can never feed a conditional or loop.
• Filters bind tighter than ?? — {{title|upper ?? "X"}} applies upper before the fallback check; parenthesize if the fallback needs the filter instead.
• {% set %} accepts selector results ({% set comments = selector:.comment %}), which pairs with bracket indexing ({{timestamps[loop.index0]}}) to walk two parallel arrays in one loop.

Properties

Properties become Obsidian frontmatter fields. Order matters: the extension writes properties to frontmatter in the order they appear in the array. If the target vault uses a YAML sort plugin (Linter, etc.), match the template's property order to the user's sort order so freshly clipped notes don't reshuffle on lint.

Each property in the array:

{
  "name": "property_name",
  "value": "{{variable|filter}}",
  "type": "text"
}

Type	Description
text	Single string value
multitext	Array (comma-separated values become list items)
date	Date value (ISO format or use date filter)
number	Numeric value
checkbox	Boolean — expects "true" or "false"

JSON Escaping for Interpreter Prompts in Properties

Interpreter prompts inside JSON property value strings require escaped quotes. The inner " that delimit the prompt must be escaped as \\\" in the JSON:

{
  "name": "description",
  "value": "{{\\\"Summarize in 1-2 sentences\\\"}}",
  "type": "text"
}

The escaping layers:

1. Outer " → JSON string delimiter
2. \\\" → produces literal \" in the parsed string
3. Template engine sees {{"Summarize in 1-2 sentences"}} and sends to interpreter

Static values and preset variables don't need this escaping — only interpreter prompts inside JSON strings.

Filters on interpreter prompts in properties also need escaping:

"value": "{{\\\"author's full name\\\"|wikilink}}"

Triggers

Auto-select this template when the URL or page schema matches.

Type	Format	Example
URL prefix	Plain URL string	"https://letterboxd.com/film/"
Regex	Enclosed in /	"/^https?:\\/\\/docs\\.google\\.com\\/forms/"
Schema type	schema:@Type	"schema:@NewsArticle"

Multiple triggers in the array act as OR — any match selects the template.

Example: Full-Text Article Template

A vendor-neutral article clipper. Captures the full article body and a small set of metadata; mixes preset variables with optional AI-interpreter prompts. Treat the folder, tag, and prompt choices as illustrative — substitute the user's own conventions before delivering.

{
  "schemaVersion": "0.1.0",
  "name": "Article",
  "behavior": "create",
  "noteNameFormat": "{{title}} - {{site}}",
  "noteContentFormat": "{{content}}",
  "path": "Clippings",
  "properties": [
    {
      "name": "title",
      "value": "{{title}}",
      "type": "text"
    },
    {
      "name": "url",
      "value": "{{url|split:\\\"?\\\"|slice:0,1}}",
      "type": "text"
    },
    {
      "name": "author",
      "value": "{{author}}",
      "type": "text"
    },
    {
      "name": "site",
      "value": "{{site}}",
      "type": "text"
    },
    {
      "name": "published",
      "value": "{{published}}",
      "type": "date"
    },
    {
      "name": "description",
      "value": "{{description}}",
      "type": "text"
    },
    {
      "name": "tags",
      "value": "{{\\\"2-3 comma-separated topical tags, lowercase, max two words each\\\"}}",
      "type": "multitext"
    },
    {
      "name": "created",
      "value": "{{date}}",
      "type": "date"
    }
  ],
  "triggers": []
}

Key patterns in this example:

• noteContentFormat: "{{content}}" captures the full article body in Markdown.
• noteNameFormat combines page title with site name; swap to whatever filename shape the user prefers.
• URL stripped of query/tracking params: {{url|split:\\\"?\\\"|slice:0,1}}.
• description uses the page's own description ({{description}}); replace with an interpreter prompt only if the user wants AI-generated summaries.
• tags uses an AI-interpreter prompt for topical tags. Drop or replace with a static value if the user doesn't have the interpreter enabled.
• Dates use preset variables directly.

For domain-specific templates (recipes, films, books, jobs, etc.), prefer schema-driven extraction ({{schema:@Type:key}}) and schema:@Type triggers — schema-based templates auto-match across many sites of the same conceptual domain.

Adapting AI Prompts into Interpreter Variables

To convert an existing AI prompt into a web clipper interpreter variable:

1. Place the prompt text inside {{"..."}} for use in noteContentFormat or noteNameFormat
2. For use inside a JSON property value, escape as {{\\\"...\\\"}}
3. Add filters after the closing quote: {{"prompt"|filter1|filter2}}
4. If the prompt should return structured data, instruct it to return JSON, then chain map and template filters
5. Keep prompts focused — the AI only sees page content (or whatever context provides)

Validating before delivery

Template parse errors are silent until the user tries to import — and the import error is often laconic ("Unexpected character 'X' in template" with a line number pointing at the JSON, not the template body). Before declaring a template done:

• Mentally walk through every map, template, and replace chain for the gotchas above.
• If the user has access to the extension, ask them to test-import the JSON before iterating further on the body or properties — a clean import is the fastest signal that filter syntax is valid.
• If iterating with the user on a working template, ship one change at a time so a reintroduced parse error is easy to attribute.

Reference

For exhaustive filter signatures, the full preset variable list, schema variable edge cases, and anything not covered above, fetch the live docs. Two source-of-truth locations:

• The clipper repo's docs/ (e.g. obsidianmd/obsidian-clipper, look for docs/Filters.md, docs/Variables.md, docs/Templates.md, docs/Logic.md, docs/Interpret web pages.md)
• The official Obsidian help docs' Web Clipper section (e.g. obsidianmd/obsidian-help)

Canonical human-facing docs: https://obsidian.md/help/web-clipper (subpages /filters, /logic, etc.). These are JS-rendered; a plain fetch returns an empty shell. Link users there, but read from GitHub.

Use whatever retrieval tool fits — context7, WebFetch, direct GitHub raw URLs, or anything else. Pick the one that's faster or more accurate in the moment. Prefer the live docs over guessing from memory.