Turn a document, chat log, braindump, or the _inbox into linked one-idea atoms plus index links, following the vault's conventions. Use when a document, a chat log or the inbox needs to go into the vault as linked notes. Activates on "/atomize", "atomize this", "sweep the inbox", "file this into the vault".
Installs into .claude/skills of the current project.
Are you the author of Atomize?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/uncovertechtalent-atomize)
---
name: atomize
description: >
Turn a document, chat log, braindump, or the _inbox into linked one-idea atoms
plus index links, following the vault's conventions. Use when a document, a
chat log or the inbox needs to go into the vault as linked notes. Activates on
"/atomize", "atomize this", "sweep the inbox", "file this into the vault".
user-invocable: true
---
# Atomize
Input: a file path, pasted text, or (no argument) the vault's `_inbox/`. Output: atoms filed in domain folders, links between them, source handled.
## Procedure
1. **Read the vault's CLAUDE.md first.** Its conventions, folder list, and private boundary override anything here.
2. **Extract ideas, not sections.** An atom is one claim, decision, fact-cluster, or reusable reference, stated in its first line. Headings in the source are formatting, not atom boundaries; a single section can hold three ideas and three sections can hold one.
3. **Search before creating.** For each candidate atom, run `vault_search` for existing coverage. Update or extend an existing atom rather than writing a near-duplicate. Genuinely new ones get new files.
4. **Name for the future reader.** Filename = the idea, sentence-cased. Not the source ("Meeting 2026-08-03 part 2"), the content ("Vendor X owns the schema migration").
5. **Link as you go.** Atoms from the same source usually relate; link them. Also link into existing atoms surfaced by the searches in step 3. Unresolved links to atoms worth writing later are encouraged.
6. **Handle the source.** Inbox items: delete after atomization (their content now lives in atoms). External files: leave in place, add a one-line pointer at the top of the source noting it was atomized and where, if the file will be seen again.
7. **Update MOCs.** If a domain crossed the ~7-atom threshold, create its `_MOC.md`; if a MOC exists, add the new atoms with one-line summaries.
8. **Reindex and report.** Run `vault_reindex`, then report: atoms created (with paths), atoms updated, links added, inbox items cleared. Counts come from the finished work, not the plan.
## Judgment calls
- Boring operational text (logistics, scheduling) usually doesn't earn atoms. When most of a source is noise, say so and extract the two things that weren't.
- Long verbatim material worth keeping (a quoted email, a spec) can live as an atom-with-a-body; the first line still states why it's kept.
- Anything touching the CLAUDE.md private boundary: file it inside the boundary and never quote it outside the vault.
## Overlay
This skill works on the generic defaults in this file. Before you start, check for a private overlay:
```bash
[ -n "$POWERSUIT_OVERLAY" ] && ls -1 "$POWERSUIT_OVERLAY/vault" 2>/dev/null || true
```
If `$POWERSUIT_OVERLAY/vault/` exists, read its `README.md` first, then the files in `vault/atomize/` if that folder exists. Read other overlay files only when that README or this skill names them. Where the overlay and this file disagree, follow the overlay. If the variable is unset or the folder is missing, use the defaults and do not ask for an overlay. If the permission settings block the command, use the defaults and say in your output that the overlay was not checked. Never copy overlay content into this plugin, into its templates or into any public repository.