AI Skill Report Card
Organizing Link Dumps
Quick Start15 / 15
Given a raw dump like:
https://react.dev/learn
cool article about css grid https://example.com/css-grid
https://github.com/facebook/react
tailwind docs: https://tailwindcss.com/docs
https://react.dev/learn (dupe)
random thought: look into bun vs node
https://nodejs.org/api/fs.html
Produce:
Markdown# Link Dump — Organized
Recommendation▾
Add an example showing handling of malformed/broken URLs to demonstrate the 'Needs Review' section in practice
Frontend Frameworks
- React Docs — official learning guide
- React (GitHub) — source repo
CSS
- CSS Grid Article — explainer article
Tooling / Docs
Notes / To Research
- Compare Bun vs Node.js
Workflow14 / 15
Progress:
- Step 1: Ingest raw dump; split into individual entries (one URL/thought per line, handle inline mixed text)
- Step 2: Extract URLs, strip tracking params (utm_*, ?si=, etc.), normalize (https, trailing slashes)
- Step 3: Deduplicate (exact URL match + near-duplicate domain+path match)
- Step 4: Classify each entry — infer category from domain, URL path, and any surrounding text
- Step 5: Pull or infer a short title/label for each link (from text context, or domain name if no title given)
- Step 6: Separate non-URL content into "Notes" or "To Research" bucket — don't discard it
- Step 7: Group into categories; merge tiny categories (1-2 items) into an "Misc" section unless clearly thematic
- Step 8: Sort within categories (alphabetical or by apparent importance — pick one, be consistent)
- Step 9: Output structured markdown with headers, bullet lists, and one-line annotations where context exists
- Step 10: Flag ambiguous/broken entries (malformed URLs, unclear intent) in a final "Needs Review" section
Recommendation▾
Clarify the near-duplicate detection logic (e.g., what threshold distinguishes a near-dupe from a distinct link)
Examples17 / 20
Example 1: Input:
youtube.com/watch?v=abc123&t=42s - react tutorial
https://news.ycombinator.com
stackoverflow.com/questions/123/how-to-center-div
buy groceries
https://youtube.com/watch?v=abc123
Output:
Markdownundefined
Recommendation▾
Consider showing a before/after on a larger, messier dump to demonstrate scalability of the categorization approach
Video
- React Tutorial (YouTube) — dedup'd, timestamp dropped
News / Aggregators
Notes / To Research
- buy groceries (not a link — personal reminder, consider moving to a task list)
**Example 2:**
Input: 40 unlabeled URLs, all GitHub repos.
Output: Single "GitHub Repos" category, grouped further by inferred purpose (e.g., "CLI Tools," "ML/AI," "Learning Resources") based on repo name/description patterns, sorted alphabetically within subgroups.
Best Practices
- Preserve every piece of original content — categorize or flag, never silently drop.
- Infer categories from evidence (domain, path keywords, surrounding text), not guesswork dressed as certainty.
- Keep annotations short (one line); don't invent detail the source didn't provide.
- When a link's purpose is unclear, label it plainly rather than fabricating a description.
- Use consistent category granularity — don't create a category for a single item unless it's clearly a distinct topic.
- Strip tracking/session params from URLs but preserve functionally necessary query params (e.g., search queries, video IDs).
Common Pitfalls
- Don't over-categorize into dozens of micro-sections — merge sparse categories.
- Don't treat visually similar URLs as duplicates without checking (e.g., different video IDs, different SO question IDs).
- Don't discard non-URL text as noise — it's often the most context-rich part of the dump.
- Don't invent titles/descriptions not supported by the link or surrounding text — use the domain name as fallback.
- Don't assume one correct category per link in ambiguous cases — pick the most useful primary category and optionally cross-reference.