Skip to content

← Back to Doc Crawl and Summarize

SKILL.md

v1.0.0 · 2203 B · text/markdown

---
name: doc-crawl-summarize
description: Use when you need up-to-date information from a documentation site, GitHub issue tracker, or arbitrary URL - fetch and convert to clean markdown instead of relying on possibly-stale training knowledge
---

# Doc Crawl & Summarize

Training knowledge about a library's API goes stale the moment it ships a
new version. This skill's discipline: when the task depends on external,
possibly-changed information, go get the current version instead of
guessing from memory.

## When to reach for this

- The task references a specific library/API version that may have changed
  since training.
- A user links a specific doc page, issue, or PR and expects the answer to
  reflect its actual current content.
- You're about to state a specific API signature, config option, or
  behavior you're not fully certain is still accurate.

## How to fetch well

1. **Fetch the specific page first**, not a guessed URL — if you don't have
   the exact URL, search for it rather than constructing one from a
   pattern that might not hold.
2. **Convert to clean markdown/text before reasoning over it** — raw HTML
   is noisy; strip navigation chrome, ads, and boilerplate so the actual
   content is what gets reasoned about.
3. **For a whole documentation site**, crawl breadth-first from the index
   page rather than guessing subpage URLs, and stop once you have the
   sections actually relevant to the task — don't crawl the entire site
   for a question about one function.
4. **For GitHub issues/PRs**, fetch the issue/PR content and its comments
   directly rather than relying on the title alone — the real answer is
   often in a comment thread, not the original post.

## After fetching

State what you found and cite where it came from (the URL), especially
when it corrects or updates something you'd otherwise have stated from
memory — the point of fetching is to be verifiably current, not just to
have looked something up.

## What this isn't for

Don't fetch pages for information that's stable and well-established
(language syntax, well-known algorithms) — reserve this for genuinely
volatile, version-specific, or source-of-truth information.