| name | source-first |
| description | Use when user runs "/source-first", asks to "check official docs", "look up the source", "read the original blog", or wants authoritative first-party answers instead of second-hand summaries. Covers any technology question where primary sources (official docs, blogs, source code, changelogs) exist. |
Source-First Research
Answer technical questions exclusively from primary sources. Never rely on memorized knowledge or second-hand summaries when this skill is active.
Source Priority
Always follow this hierarchy — stop at the first level that answers the question:
- Official documentation — resolve via context7, then fetch
- Official blog posts / changelogs — fetch the URL with defuddle
- GitHub source code / README / releases — read via GitHub MCP or
gh CLI
- Web search as last resort — tavily/web search, but only for locating a primary URL to then fetch directly
Never present search engine snippets as the answer. Search is only for finding the primary URL, then fetch and read the actual page.
Tool Selection
| Source type | Tool | When to use |
|---|
| Library/framework docs | context7 (resolve-library-id then query-docs) | Question about a specific library's API, usage, or features |
| Blog post / announcement | defuddle (obsidian skill) | User provides URL, or you locate an official blog URL |
| Web page (fallback) | WebFetch | When defuddle is unavailable or fails |
| GitHub repo / source | gh CLI or GitHub MCP tools | Source code, README, releases, issues |
| Locating URLs | tavily-search or WebSearch | Only to find the official URL, then fetch it with the tools above |
context7 workflow
1. resolve-library-id with library name
2. query-docs with the resolved ID and the specific question
3. Extract relevant sections from the response
defuddle workflow (preferred for web pages)
1. Obtain the official URL (from user, or from a search)
2. Use defuddle to extract clean markdown content
3. If defuddle fails, fall back to WebFetch
GitHub workflow
1. gh repo view owner/repo (overview)
2. gh api repos/owner/repo/releases/latest (changelogs)
3. Read specific files: gh api repos/owner/repo/contents/path
Output Format
Structure every response as:
## Source
[URL or source identifier]
Retrieved: [timestamp or "from context7 docs"]
## Key Excerpt
> [Direct quote or extracted content from the primary source]
> [Include enough context to be self-contained]
## Analysis
[Your interpretation, summary, and answer to the user's question]
[Connect multiple sources if needed]
[Flag if the source is dated or potentially outdated]
Rules
- No memory-based answers. If you cannot find a primary source, say so explicitly rather than answering from training data.
- Quote before interpreting. Always show the original text first, then your analysis.
- Attribute everything. Every claim must trace back to a specific source with a URL or file path.
- Prefer narrow fetches. Use context7 topic queries or grep-style searches over fetching entire pages when possible.
- Admit gaps. If the primary source doesn't fully answer the question, state what it covers and what remains unclear.
