← ClaudeAtlas

ingest-alllisted

Compile every uncompiled source in raw/inbox in one pass — plan the whole batch first, approve once, then write. Use when the user says "ingest all", "compile everything", "clear the inbox", or when several sources are waiting and compiling them one at a time is just friction.
playingpossumm/loose-ends · ★ 1 · AI & Automation · score 74
Install: claude install-skill playingpossumm/loose-ends
# ingest-all Same work as `/ingest`, done across the whole inbox with **one approval instead of one per source**. Everything in [`../ingest/SKILL.md`](../ingest/SKILL.md) still applies — the rules about citation, provenance, contradictions and loops are not relaxed because there are more files. What changes is only where the human checkpoint sits: once, over the whole batch, instead of between every source. Read [CLAUDE.md](../../../CLAUDE.md) first if you have not this session. ## Drain first Before planning anything, run: ``` python scripts/telegram_capture.py --once ``` `/ingest-all` compiles what is in `raw/inbox/`, and Telegram messages only land there when that script runs. Without this step, "compile everything" silently means "compile everything that arrived before the last scheduled drain" — so a note sent from your phone an hour ago is missed, and the run reports success anyway. `--weekly` already drains before writing the brief, for the same reason. Draining twice costs nothing: the script tracks an offset, so a second run in the same minute fetches nothing. ## What writes itself, and what waits Not every source carries the same risk, so not every source needs the same approval. Plan the whole batch first, as always, then split it. The test is **which store the plan writes to**, which is the two-store distinction in `CLAUDE.md` applied one step earlier — at the decision of whether to write, not only at what may be written. | The plan writes | Stakes | T