Skip to Content
V1GovernanceMandatory Rules

Mandatory, linked and available Rules

Not every Rule belongs at the top of every AI session. Lore delivers a Rule in one of three ways, depending on what your org decided and how the Rule was published.

How it reaches a sessionDecided by
MandatoryAt every session start (governance hooks) and in Lore MCP responsesThe org’s policy
LinkedTogether with the Skill or package that depends on itThe package’s dependencies
AvailableFound and loaded on demand (lore_search, lore_get, lore pull)Nobody: it’s just published

The Spec-Driven Development framework your projects follow is a separate, single-choice policy: see SDD frameworks.

Mandatory: imposed by the org

Imposing a Rule is your org’s decision, not a property of the package. The same public Rule can be mandatory in one org and optional in another. The org owner lists the mandatory packages in the painel, under the org’s Mandatory Packages page. Each entry is:

  • a Rule or a Skill, as namespace/name;
  • optionally a version constraint (latest by default);
  • optionally project globs, matched against the repository’s git remote (for example dev.azure.com/acme/web/*), to impose it only on some repositories.

What each session receives:

  • Rules arrive with their full text (up to 1500 bytes; a longer Rule arrives as its description and a pointer to lore_get);
  • Skills arrive as a one-line listing, loaded with lore_get when relevant, because a Skill is too large for every session.

Project matching happens on your machine: the hook never sends your repository’s remote to Lore for it. MCP-only surfaces, such as claude.ai and ChatGPT, have no repository, so they receive only entries without project globs.

On MCP-only surfaces

claude.ai, Claude Desktop and ChatGPT have no session-start hook, so the Rules reach them only through the Lore MCP connector:

  • ChatGPT: a connector scoped to your org (?org=<slug>&auth=required) tells the model, when it connects, to call lore_context at the start of each conversation. That call returns the org’s current Rules. ChatGPT keeps a copy of this instruction until the connector is refreshed, so refresh the connector after changing its URL. Policy changes need no refresh: the instruction never contains the Rules themselves.
  • claude.ai and Claude Desktop: they don’t pass connector instructions to the model. On their own, the Rules arrive only when the model calls a Lore tool, appended to that tool’s result, which is best-effort. To cover every conversation, an org owner copies the text under claude.ai organization instructions on the painel’s Mandatory packages page and pastes it in claude.ai’s organization settings , under Organization and access → Organization instructions (Team and Enterprise owners; up to an hour to apply). The text tells the model to call lore_context and summarizes the unscoped Rules as a fallback. Lore can’t set it for you, because claude.ai has no API for it. The painel shows who copied it last and warns when the policy changed after that copy.

The owner also sees who received the current policy, per member, through the governance hook and the MCP connector.

An org that never configured the policy keeps the original behavior: every Rule in the org is delivered, as a one-line description. Configure starts from that list, so the owner only removes what isn’t imposed.

Linked: brought by a package

A Rule that only matters for one Skill, framework or package should be a dependency of that package, in its manifest.yaml:

dependencies: - acme/nextjs-conventions:^1.0

It then arrives whenever that package is loaded (lore_get includes dependencies by default, and lore pull installs them), and never clutters sessions that don’t use it.

Available: everything else

Any other Rule your org can read stays one search away. The AI can find it with lore_search and load it with lore_get, and developers can install it with lore pull.

© 2025 Lore · Built by M1 Cloud