Capture

Capture

Write a note into the vault from a source the user hands you. One arc, five steps: search → read the primary source → create through sb → leave it in the inbox → offer connections.

This is the dominant vault interaction. Do not decompose it into separate asks; the user saying "read this and make a note" means all five steps, and the connect step in particular will not happen unless this skill runs it.

See references/note-format.md for frontmatter and body shape. See ../obsidian/references/sb-cli.md for the full sb command surface.

Step 0: Locate the vault

npx @techpickles/sb vault info
npx @techpickles/sb config qmd-collection

Read the vault's own CLAUDE.md if you haven't this session. Vault-specific policy (destinations, status vocabulary, disambiguation rules) overrides the defaults here.

Step 1: Search before creating - never skip this

Run a lex + vec pair through the qmd MCP tools, not the shell. (mcp__qmd__query is the working path; shelling out to qmd hits Metal-init failures under the sandbox.)

mcp__qmd__query
  searches: [{type: 'lex', query: '<topic terms>'},
             {type: 'vec', query: '<what the note would say>'}]
  collection: <from sb config qmd-collection>
  intent: '<why you are searching>'
  rerank: false        # CPU-only box, reranking is not worth the latency here

Show the user what already exists - path + score + snippet - before anything gets written. Often the right move is extending a note, not minting one. If a result is a strong match, say so and ask which way to go instead of quietly creating a duplicate.

When search comes up empty

A miss is not proof the note is absent. In order:

1. Basename is the stable identity. Notes get routed between folders constantly. Search the basename across the whole vault (Glob/Grep) before concluding anything is missing.
2. Check both inboxes. 10-19 System & Capture/11 Inbox & triage/ and 10-19 System & Capture/11 Inbox/ are separate folders with the same JD code. sb inbox list only reads the first.
3. The note may genuinely not exist. Say that plainly and offer to start one.

When the user says "that isn't the one"

Do not re-run the same query with different phrasing, and do not widen -n and hope. Ask for one distinguishing detail (who made it, where they saw it, what it sits next to) and search on that instead. Re-running near-identical searches has burned entire sessions, including the same lookup failing twice across two separate sessions.

Step 2: Read the actual primary source

Never write a note from search snippets. Fetch the real thing:

If you cannot reach the source, or cannot find the thing the user described, stop and say so. Do not write a note containing a URL you did not load, a title you inferred, or details you filled in from background knowledge. Reporting "I couldn't find anything matching this" is the correct outcome and is more useful than a plausible note that is wrong.

Step 3: Create the note through sb

npx @techpickles/sb note create --source auto --title "<title>" --dry-run   # check the path first
npx @techpickles/sb note create --source auto --title "<title>"            # returns JSON with path

Then write the body at the returned path with Write/Edit.

Do not hand-roll the destination path. sb resolves the inbox, the zettelkasten timestamp, and the filename slug (202608141249 note-title.md - space after the timestamp, hyphens inside the slug). Hand-rolled Write calls are how captures end up in the inbox the pipeline never reads: as of 2026-08-14 that had orphaned 46 notes, and untangling it took a single 211-message session.

For a title, prefer the source's own framing over a topic label. # Glide browser beats # Notes on a keyboard-driven browser.

Step 4: Leave it in the inbox for process-inbox to route

The note stays where sb put it. Routing is process-inbox's job, not capture's — don't ask, don't route it, and don't mention that it still needs routing.

This overrides the vault's general "capture: file rough, move on / a wrong-but-close home beats the global inbox" guidance, which governs notes the user files by hand. Notes the agent writes on request stay in the inbox until process-inbox picks them up. Both rules coexist in the vault CLAUDE.md; the scope of the override has historically been read as ambiguous — first producing a roughly 50/50 split between inbox and direct-to-area filing, later a per-note routing question duplicating process-inbox's own batch routing. It is not ambiguous: agent-written note on request → inbox, no offer, no mention.

The exception is an explicit destination named in the same conversation ("put it in 66"). Resolve it the way route would rather than hand-rolling a path: run sb vault structure, match the user's phrase against a real destination (JD area/ code or PARA folder), then sb note move --from "{note-path}" --to "{destination}/". Never construct the destination path from the user's words alone.

Step 5: Offer connections, append the daily breadcrumb

Connection discovery does not fire on its own, and the user should not have to notice it was skipped. Offer it every time, using the step 1 results plus a fresh search from the finished note's content.

• Only link notes confirmed to exist by an actual search hit. Never emit a [[link]] to a note you have not verified, and never create side notes on your own initiative. Surface those as suggestions ("want me to make a note for X too?") and let the user decide.
• Add links under ## Related with a few words on why each one relates. A bare link list decays.

Then append the daily-note breadcrumb yourself - this is no longer a yes/no:

1. Ensure today's daily note exists. sb daily append ENOENTs on a missing daily note instead of creating one. Run npx @techpickles/sb daily path to get today's path, then check it with Read/Glob:
   • If missing, read .obsidian/daily-notes.json for the template path, Read that template, and Write it verbatim to today's path before appending. If no template is configured, say so and stop rather than guessing a daily-note shape.
   • If the append still fails for any other reason, say so out loud. A dropped breadcrumb must never be silent now that no one is asked before it happens.
2. Append:

   npx @techpickles/sb daily append --section "Notes" --content "- [[<note>]] - <one-line why>"
   

   Keep it to links unless asked for more. Do not restructure or rewrite the daily note.
3. Verify it landed under ## Notes, with a temporary fallback. sb daily append has a live bug where --section fails to match an existing H2 and dumps the line as a stray plain-text block elsewhere in the file instead of under the target section. Read the daily note back and confirm the new line sits under the real ## Notes heading. If it landed somewhere else, remove the stray block and Edit the line into the correct ## Notes section yourself. Delete this verification step once the CLI's section matcher is fixed - it exists only to paper over that bug, not because this is the intended design.
4. Report it in one line, e.g. Added breadcrumb to today's daily note. No ceremony - just enough for Josh to see it and undo or adjust.

For deeper backlink weaving, hand off to the connect skill rather than reimplementing it.

Capturing several sources at once

When the user hands over a batch ("make notes for each of these", "write them up"), dispatch one subagent per note so each reads its source in isolation and cannot cross-contaminate. This is a pattern the user asks for explicitly; default to it for 3+ sources.

Each subagent: read the primary source, write one note via sb, and report back path, title, verified wiki-links used, and anything it could not confirm. Then do the cross-linking pass yourself once all notes exist, since connections need the whole set in view.

Constraints

• Search first, always. No note gets written before the user has seen what already exists.
• Primary sources only. No note written from a snippet, a memory, or a guess.
• Never fabricate. No invented URLs, no unverified [[links]], no filled-in details. Stop and report the gap instead.
• sb owns paths. No hand-rolled destinations.
• Inbox by default. Routing is process-inbox's job — capture never asks or routes, except an explicit user-named destination, resolved via sb vault structure like route does, never hand-rolled.
• Connecting is offered; the breadcrumb is not. Suggest related links and let the user decide, but the daily-note breadcrumb append happens automatically every time - it no longer waits for a yes.
• One note, one idea. If the source carries several distinct ideas, say so and offer to split rather than writing one sprawling note.