Bulk-import many sources at once

Read the note format and vault layout reference and capture's SKILL.md first — this skill reuses capture's per-item write rules almost unchanged and only adds a batch wrapper around them. Also read defuddle's safety contract if the batch has any URLs.

This skill files many sources in one pass. It does not run automatically — only when the user hands Claude more than one source to file: a directory ("import everything in this folder"), an inline list of paths/URLs in the request text, or a plain text/Markdown file where each line is a path or URL. A single source is capture's job, not this skill's; a research question to investigate (rather than a list of sources to file as-is) is autoresearch's job.

Steps

1. Resolve the vault per Resolving the vault, the same way capture does. Create inbox/, notes/, projects/, indexes/, templates/, and attachments/ in it if any don't already exist.

2. Build the source list:

   • Directory: list its immediate contents only — one level, no recursion into subdirectories, matching the vault's own shallow- nesting rule. A nested directory the user clearly means to include should be named explicitly in a separate request; don't expand it automatically.
   • Inline list: each path or URL the user wrote directly in the request text is one source.
   • Listing file: read the file; each non-blank line is one source (strip a leading -/* bullet marker and surrounding whitespace if present).

   Classify each entry: starts with https:// → URL; starts with http:// or otherwise looks like a URL but isn't https:// → still a URL, which the safety contract in step 3 will reject; anything else → local path.

3. If the batch includes any URLs, before fetching anything:
   1. Validate each URL against defuddle's safety contract (HTTPS only; no embedded credentials; no private/local hosts; no sensitive query parameters; etc.). Drop any that fail validation — record each as skipped (skipped: rejected — <reason>) in this run's report; a rejected URL never appears in the consent list below.
   2. Check whether a reviewed Defuddle-style extractor is available — the same one-time check defuddle itself does:
      • Not found: skip every remaining URL source for this run (each reported as skipped: no extractor) and continue to step 4 with only the local sources. Don't block a batch that's mostly local files on one missing external tool.
      • Found, unreviewed: show its resolved path and get the user to confirm its provenance and version once, for the whole batch — not once per URL.
      • Found and already reviewed this session: continue.
   3. If any URLs survived validation and an extractor is available, get one upfront batch consent: list every surviving URL and its host, state plainly that all of them will leave the machine, and stop for explicit approval before fetching any of them. This single approval covers every URL in the list — no per-source consent inside an already-approved batch.
4. Process each source in sequence — not in parallel:
   • Local source: apply capture's steps 3, 4, 5, 6, 7, 8, and 10 (read/view the source, copy a binary attachment, derive a title/slug, decide the destination, write the note, update the topic index if it landed in notes/<topic>/, and the linking pass) to this one source. Skip capture's step 9 (regenerating the master index) — this skill regenerates it once, after the whole batch, in step 5 below.
   • URL source: apply defuddle's fetch-and-clean step (its redirect policy and fail-closed rules — already covered by this run's batch consent), then the same capture steps as the local-source case above, using defuddle's cleaned Markdown as the source content.
   • Bulk-specific overrides to those per-item rules (apply instead of the ask-first behavior capture's steps 5 and 6 normally use):
     • A filename collision (capture step 5) is skipped, not asked: record it in this run's report (skipped: collision with <existing path>) and move to the next source. The user resolves it afterward with a normal capture or a manual edit.
     • Genuinely ambiguous project-vs-note placement (capture step 6) defaults to inbox/, not asked: file it there and record it as filed-to-inbox in the report. The user triages inbox/ afterward with organize.
     • Any other per-item failure — unreadable file, empty or failed fetch, a defuddle extraction error — is recorded (skipped: <reason>) and the batch keeps going.
5. After the whole batch finishes, regenerate indexes/README.md once per Master index — one full rescan at the end covers every change the batch made, so there's no need to regenerate per item.
6. Report a per-source summary, grouped into: filed (path, and which topic index was updated, if any), filed to inbox/ (ambiguous placement), and skipped (with each one's specific reason) — so the user can see exactly what happened across the whole batch and what still needs attention.

Out of scope

No real fan-out/parallelism — sources are processed one at a time, in order; see the design spec's "Rejected: real fan-out parallelism" for why. No recursive directory expansion. No new ledger, transaction bundle, or persistent run log — the end-of-batch report in the conversation is the only record, same as every other skill in this project.

Error handling

• Missing or unreadable source, empty/failed fetch, or an extraction error: record as skipped with the reason, don't stop the batch.
• Filename collision: skip, don't ask — see step 4's override above.
• Ambiguous project-vs-note placement: default to inbox/, don't ask — see step 4's override above.
• No reviewed Defuddle-style extractor available: skip all URL sources (skipped: no extractor), keep going with local sources.
• Vault directory not writable: report the error and stop the whole batch (same as capture).