Ingest

Sequential processing (1-2 files)

For each new file:

1. Identify file type and read references/file-types.md for type-specific handling
2. Read the source thoroughly — for PDFs, use chunked reading (pages parameter)
3. View images — if source contains images, read text first, then view images for additional context
4. Identify concepts — list all concepts, entities, and topics in the source
5. Check existing coverage — for each concept, check if a wiki article already exists (use /wikillm:obsidian-cli search or Grep)
6. Decide action — create new article / update existing article / skip (already well-covered)
7. Write/update articles following the Article Writing Guide below
8. Cross-link all related articles with [[wikilinks]]
9. Write summary page for the source itself
10. Update indices — INDEX.md, TAGS.md, SOURCES.md, RECENT.md
11. Append to LOG.md
12. Git add, commit, push — one commit per source file

Parallel processing (3+ files)

Phase 1: Pre-scan (orchestrator)

The orchestrator reads and analyzes all new sources before dispatching. This is where the intelligence lives.

1a. Inventory each source:

For each new file, determine:

• File type — markdown, PDF, HTML clip, YouTube transcript, code, etc. (read references/file-types.md for handling notes)
• Size/complexity — page count for PDFs, line count for text, number of images
• Topic scope — broad survey vs deep dive on a single topic
• Key concepts — skim the source (intro, headings, abstract, first ~100 lines) to extract the main concepts, entities, and topics

1b. Read current wiki state:

• Read wiki/_index/SOURCES.md and wiki/_index/INDEX.md to know what already exists
• Identify existing articles that new sources might update

1c. Build concept assignment table:

• List every concept across all sources
• If a concept appears in multiple sources, assign ONE source as the owner (the source that covers it most deeply)
• Other sources get the concept marked as link-only (reference it, don't create/update the article)
• Existing wiki articles that multiple sources touch: assign one owner the same way

1d. Decide dispatch strategy:

Use your judgement based on the inventory. The goal is: don't spawn subagents when the orchestrator can handle it faster itself.

• Do it yourself when the total workload is light — e.g., 1-2 short articles, a handful of clips on familiar topics, or sources that mostly update existing articles rather than creating new ones. Subagent overhead (startup, prompt injection, reconciliation) isn't worth it for small batches.
• Dispatch parallel workers when there are 3+ substantive sources, or fewer sources that are each large/complex enough to justify isolation (e.g., two 100-page PDFs).
• For very large sources (200+ page PDFs), note in the worker prompt so it can chunk appropriately.

This prevents duplicate articles and gives each worker clear boundaries.

Phase 2: Dispatch workers

Dispatch one ingest-worker subagent per source file, all in parallel, each with:

• The source file path
• The current SOURCES.md + INDEX.md content
• The concept assignment table: which concepts this worker owns (create/update) vs link-only (just use [[wikilinks]])

Phase 3: Reconcile (orchestrator)

After ALL workers complete:

1. Handle failures — if a worker errored out, check what it partially wrote by grepping wiki/ for files with that source in frontmatter sources: field. Either re-run that single source sequentially, or accept partial results and log the gap.
2. Dedup check — scan wiki/ for any articles created by multiple workers (same filename or near-identical titles). Merge if found.
3. Cross-link audit — check that worker-created articles link to each other where relevant. Workers can't know about sibling workers' new articles, so add missing [[wikilinks]] now.
4. Update indices — INDEX.md, TAGS.md, SOURCES.md, RECENT.md from aggregated worker output (including partial results from failed workers)
5. Append to LOG.md — one entry per source, mark failures with [partial]
6. Git commit + push — one batch commit or one per source file

3. Article Writing Guide

Structure

---