second-brain: Command for Claude Code

Your job is to turn new raw captures into clean, linked wiki notes.

Steps:

1. Run python scripts/ingest.py list-new to find raw files not yet referenced by any note. Also list any images or PDFs in wiki/raw/ that don't yet have a companion .md file — those need to be processed too.
2. Non-text captures first. For every image or PDF in raw/ without a companion .md:
   • Read the file directly (Claude can see images and PDFs natively).
   • Create raw/<same-name>.md with frontmatter, a description, a full transcription of any visible text/tables, and an image: or pdf: reference line.
   • Do NOT modify or delete the original binary file.
   • Treat the new companion .md as a raw capture and continue with the normal flow below.
3. For each new raw markdown file (including companions just created): a. Read it according to its length. First check line count (wc -l <path> or the file header from Read). Then:
   • Under 500 lines — read the entire file in one pass. No shortcuts.
   • 500–2000 lines — read the entire file, in sequential chunks if needed. Distill aggressively into short notes, but do not skip sections.
   • Over 2000 lines — read the first ~200 lines to get structure and thesis, scan the table of contents / headers, then read in full the sections most relevant for distillation and linking. It's fine to skip appendices, repetitive examples, or sections clearly tangential to the core concepts.
   • Always report read-coverage in your final summary as read X/Y lines of <file> so shallow reads are never silent. If coverage is under 100%, note which sections you skipped and why, so the user can request a follow-up deep pass. b. Run python scripts/ingest.py stamp <path> if it has no frontmatter yet. c. Identify the 1–3 core concepts in the file. d. For each concept, decide: does a matching note already exist in wiki/notes/? Use python scripts/search.py "<concept>" to check. e. If yes → append a distilled paragraph to that note and cite the raw file (see raw/<filename>). f. If no → create a new note in wiki/notes/<slug>.md with proper frontmatter, a short summary, and [[wikilinks]] to any related existing notes. If the source was an image or PDF, embed it inline with ![](../raw/<name>.png). g. Tag framework notes. If the note captures a reusable mental model / analytical lens / decision tool you'd apply across many topics (e.g. strategy-vs-planning, revenue-models) — not just facts about one thing — make framework its first tag so it joins the framework library (queryable via the tag). This is only for knowledge frameworks, which stay searchable. Production frameworks (templates / rubrics / playbooks that a content skill applies when generating — e.g. the SNS hook playbook) do NOT belong in wiki/notes/; they live in .claude/skills/_shared/ so they stay out of semantic search and are referenced by skills.
4. Add backlinks: if note A now references note B, make sure B has an entry pointing back to A when it's meaningful.
5. Never modify or delete original files in wiki/raw/ (companion .md files you just created are fine to edit).
6. Record each processed raw file in the ledger. After you finish distilling a raw file, run:

   python scripts/ingest.py record <raw-filename> <note1,note2,...>
   

   where the second argument is a comma-separated list of note IDs (filenames without .md) produced or updated from that raw file. This is how the ledger tracks what's been ingested — without it, the file will show up as "new" on every run.
7. At the end, print a short summary: how many raw files processed, how many notes created, how many notes updated, and which decisions you're least sure about so the user can correct you.

Remember: small, focused notes are better than big ones. Split when in doubt.