unsource
Skill0 starsChanged 33 days agoMITView on GitHub
Remove a source from the vault and revert its influence across every page it touched, which is decompilation. Use when a source turns out to be wrong, unreliable, or must be deleted, or when the user says "remove that source", "that article was wrong", "undo that ingest".
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 38
4. Wait for confirmation: Never proceed automatically.
- Very specific: line 46
Remove deleted pages from index.md.
- Distinctive: line 63
Never silently drop a claim that other sources also support: Check before removing.
- Distinctive: line 66
Prefer marking over destroying when the user's intent is ambiguous, and say what you did so they can escalate.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
unsource
Compilation spreads one source across many pages. A bad source touches fifteen pages before anyone notices, and none of the published write-ups of this pattern offer a way back. This is the way back.
It is also the deletion mechanism, since removing information about a person means removing every page their source touched, which is the same operation.
Why not just git revert
Because later good edits sit on top of the bad ones. Reverting the commit throws away
everything compiled since. This works claim-by-claim instead, which is why CLAUDE.md
requires every page to record the source ids that shaped it.
Procedure
1. Identify. Resolve the user's description to a sources/ page and its raw id. If
ambiguous, list candidates and ask, because unsourcing the wrong thing is expensive.
2. Find the blast radius. grep the vault for the source id. Every page with it in
sources:, every citation pointing at it, every loop extracted from it. Read log.md for
the original ingest entry.
3. Report before touching anything:
- pages that exist only because of this source → will be deleted
- pages where it is one of several sources → claims removed, page survives
- loops extracted from it → will be closed as
killed - claims that lose their only citation → flag these individually and ask. A claim you still believe but can no longer cite is the genuinely hard case, and it is the user's call whether to drop it or find a new source.
4. Wait for confirmation. Never proceed automatically. Show the full list first.
5. Execute.
- Delete single-source pages.
- On shared pages: remove claims citing only this source, strip the id from
sources:, bumpupdated:, and leave a note in the body recording that a source was removed and when. Do not silently rewrite history. - Close its loops as
status: killedwith a reason. - Remove deleted pages from
index.md. - Move the raw file out of
raw/only if the user asked for deletion; otherwise mark itstatus: unsourcedand leave it. Removing a claim and destroying evidence are different requests.
6. Re-lint. Unsourcing creates orphans and broken links. Run /lint and fix what it
finds mechanically.
7. Log:
## [YYYY-MM-DD] unsource | <title> — N pages edited, N deleted, N loops killedRules
- Never proceed without explicit confirmation of the blast radius.
- Never silently drop a claim that other sources also support. Check before removing.
- Leave a trace. A page that lost content records that it did. Invisible deletion is how a knowledge base starts quietly lying.
- Prefer marking over destroying when the user's intent is ambiguous, and say what you did so they can escalate.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

