When not to make a skill
The best governance artefact in the estate is not a rule that was written down in advance. It is a decision, with reasons, to refuse a skill — recorded in the sg.llm capability brief.
“No — and deliberately.… A separatesg.llmskill would compete for triggering withvault-html-appon the same intent… and split the authoring contract across two files. Skills earn their keep when they cover a distinct task; this is one capability inside an existing one.”
The three rules
Two skills matching one intent is worse than one skill, because the dispatcher must now guess which one the caller meant. An overlapping skill does not add coverage — it adds ambiguity.
If a canonical contract already exists (an API reference, a schema, another skill’s authoring rules), a new skill points at it — convention D on the authoring page — rather than restating it. A restated contract drifts from the original the same way the two copies of use-sg-playwright drifted from each other: nobody edits both copies at once, so eventually only one of them is right.
The test the brief applies to itself: is this one capability inside an existing skill, or genuinely a distinct task? sg.llm failed that test against vault-html-app and was refused. A skill that passes the test is not competing with anything; a skill that fails it is a duplicate wearing a new name.
Why this belongs on the front page of the argument
Most of what this site publishes is designed but unbuilt — a critique with a direction, not a shipped mechanism. This is the opposite: a real decision, made under real constraints, that happened to generalise into exactly the governance the thesis calls for. It cost nothing to write down because it had already been thought through once, for one skill. The catalogue is what following these rules for eight skills looks like; the divergent duplicate is what it looks like when a related discipline — a single canonical copy — is missing.
Who arbitrates triggering conflicts?
The governance rules say skills compete for triggering — but with no registry, nothing today detects that two skills claim one intent. The catalogue could: a build-time overlap check on trigger vocabularies is feasible with what already exists, and has not been run.
Open. See the full list of open questions.