Look up before producing another copy
Verify the project first, then look up existing candidates by name, version, content hash and intended use before generating, downloading or copying resources. A matching filename or hash does not establish ownership or permission.
Each project keeps its library inside its own main folder. Similar subjects, names or bytes do not authorize a shared library across different projects.
Choose a form for the consumer
Read-only views let several names refer to one pinned content object. Materialize independent files when a consumer needs to edit them, and export standalone files when the result must travel without the library.
The tested copy-on-write backend uses macOS/APFS. Consult the documented commands on other storage; explicitly allowed ordinary copies incur copy cost and are not evidence of identical behavior on every platform.
Install and check dependencies
The repository root SKILL.md is the skill entrypoint. Install the complete directory through the agent’s skill installer, or run the CLI from source. Python 3.9 or newer is required; pool operations use the standard library.
Exact duplicate scans and batch plans additionally require fclones. The scripts do not install dependencies automatically. Follow the public README for installation, plan review and export commands.
Boundaries and evidence
Originals stay in place. There is no automatic deletion, background watcher, operating-system write interception or automatic cross-project sharing. Exclude libraries and views from releases, source-only copies and recovery snapshots.
A reference view depends on its library and is not an independent backup. The documented copy-cost comparison is not disk space released or a guaranteed saving for a different dataset.