From summarizer
Summarizes web content by fetching URLs, extracting quote-grounded key passages, and structuring output for documentation, articles, API references, READMEs. Reports partial accessibility.
npx claudepluginhub jamie-bitflight/claude_skills --plugin summarizerThis skill uses the workspace's default tool permissions.
Apply this methodology when summarizing web content from URLs.
Extracts clean Markdown from web pages by stripping navigation, ads, sidebars, footers, and boilerplate using Defuddle. Use for reading docs, articles, blog posts, research papers, or release notes.
Fetches URLs and extracts main content (title, description, paragraphs) with 80-99% HTML size reduction via smart extraction. Use for ad-hoc web research and quick page previews.
Routes summarization requests for files, URLs, images, and multi-source content to type-specific strategies, enforcing anti-hallucination and fidelity rules.
Share bugs, ideas, or general feedback.
Apply this methodology when summarizing web content from URLs.
The model MUST use this skill when:
The model MUST follow this decision tree for fetching URL content:
URL PROVIDED
│
├─ Is it docs.anthropic.com or code.claude.com?
│ ├─ Yes → Append .md to path
│ │ Example: /docs/guide → /docs/guide.md
│ │ Use mcp__Ref__ref_read_url
│ └─ No ↓
│
├─ Is it documentation (docs subdomain, /docs/, /api/, /reference/)?
│ ├─ Yes → Use mcp__Ref__ref_read_url
│ │ Tool optimized for documentation sites
│ └─ No ↓
│
├─ Is it a generic web page?
│ └─ Yes → Use WebFetch
│ Fall back if mcp__Ref fails
When fetching fails, the model MUST report explicitly:
Unable to access [URL]: [error details]
Reason: [HTTP 404 | timeout | SSL error | authentication required | etc.]
The model MUST NOT:
When only some sections of a page load:
## What Was Found
Content extracted from sections: Introduction, API Reference, Authentication
Total sections accessible: 3 of 7
## What Was NOT Found
Unable to access sections: Rate Limits, Error Codes, Webhooks, Changelog
Reason: Pagination links did not resolve, or content blocked by JavaScript
Adapt summarization approach based on web content type:
| Content Type | Extract | Emphasis |
|---|---|---|
| Documentation | API structure, authentication, rate limits, key concepts, version info | Technical accuracy, preserve parameter details, note external references |
| Articles/Blogs | Thesis, supporting points, evidence, conclusions, publication date | Logical flow, distinguish opinion from facts, note if dated |
| API Reference | Base URL, auth method, endpoints/methods, request/response formats, rate limits, error codes | Machine-parseable structure, exact paths, required vs optional parameters |
| GitHub README | Purpose, installation, usage examples, dependencies, license | Actionable setup steps, stable vs experimental features |
| Generic Pages | Visible text, headings, significant quotes | Report low confidence, note navigation vs content density |
The model MUST use this two-phase approach for text-heavy sources.
Before writing any summary, extract relevant quotes:
EXTRACTED PASSAGES:
1. "The API uses JWT tokens for authentication. Tokens expire after 24 hours."
(Source: Section 2.1, Authentication)
2. "Rate limit is 100 requests per minute per API key. Burst allowance: 150 req/min for 10 seconds."
(Source: Section 4.3, Rate Limiting)
3. "Webhooks are not currently supported. Feature planned for Q2 2026."
(Source: Roadmap section, footer)
Write the summary by organizing and condensing the extracted passages. Every claim in the summary MUST trace back to an extract.
If a claim cannot be traced to an extract, it is potentially hallucinated and MUST be removed or verified against the source.
SOURCE: Grounding technique from Anthropic prompt engineering documentation (https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/long-context-tips, accessed 2026-02-06).
The model MUST use the format defined in Structured Summary.
Required components:
The model MUST follow all rules in Fidelity Rules:
../summarizer/templates/{format_id}.md (default: structured). The template defines the schema, required sections, and fidelity constraints for the selected format.The model MUST NOT:
See Structured Summary for complete specification. Brief example:
---
source_type: url
source_path: "https://example.com/docs/api-v2"
method: hybrid
confidence: high
confidence_notes: "Full page accessible, structured technical reference"
---
Summary leads with most important information. What Was Found lists discoveries with source references. What Was NOT Found distinguishes absence from nonexistence. Uncertain section captures ambiguous content. Sources section includes full URL with access date.