Skill: /source-mine [source]

Turn an account, corpus, or directory you already own into attributed vault notes — the sweep that gets the graph from empty to useful, and the one nobody wants to hand-write twice.

Don't: write a note without its source and mine date in the body; an unattributed claim is unverifiable a month later and indistinguishable from something you made up. Don't mine and synthesize in one pass — extract first, decide what deserves a note second, or the interesting-but-irrelevant all gets written up. Don't touch the source: mining is read-only, always.

Inputs

source is one of: github, filesystem, google, chat-history (ChatGPT / Claude.ai / Perplexity), or vault (re-reading for unlinked material). If ambiguous, ask which — the extraction rules differ per source.

Steps

1. Read source-recipes.md for this source's endpoints, extraction targets, and known pitfalls.
2. Inventory before extracting. List what exists — repo count, thread count, directory listing — and report the scale. A sweep whose size is unknown at the start silently becomes a sample.
3. Extract to a scratch list, not to the vault. One line per candidate: what it is, why it might matter, where it came from.
4. Check each candidate against the vault before writing: obsidian search query='<candidate>' limit=5. Existing note → append an ## Update — <date> section. No note → create.
5. Write the notes. Route by type:
   • a durable idea or framework → Knowledge/Concepts/ via /ingest (it enforces the Vault Schema)
   • project state → Knowledge/Projects/<project>/
   • an external artifact worth keeping → Knowledge/Reference/
   • anything you cannot place → Inbox/, for /inbox-triage to route
6. Attribute in the body: *Source: <source> — mined <date>*.
7. Commit the vault once at the end, not per note:

   VAULT="$HOME/Library/Mobile Documents/iCloud~md~obsidian/Documents/Personal"
   git -C "$VAULT" add '<the note paths you wrote>' && git -C "$VAULT" commit -m "docs: mine <source> → N notes"

Output

Report, before the commit:

Source: <source>          Inventoried: <n>    Extracted: <n>    Written: <n>
Created:  <paths>
Updated:  <paths>
Skipped:  <n> — <why: transient, duplicate, out of scope>
Not read: <n> — <why: timeboxed, auth failed>

Skipped and Not read are the load-bearing lines. Without them a partial sweep reads as complete.

Scope guard

Mining reads widely across personal accounts. Extract only what serves the stated goal — do not compile a profile because the data is reachable. When a source needs a sign-in that isn't already active, stop and ask rather than working around it.