The three priorities
| Priority | Meaning |
|---|---|
must | A binding constraint. The agent follows it even when it conflicts with a general best practice or its own default behavior. |
should | A strong default. The agent applies it unless there is a clear, stated reason to deviate. |
context | Reference knowledge — facts, contacts, glossary terms an agent looks up when relevant. Not a rule to follow, and not pushed into the pack. |
How priority shapes the bootstrap pack
The pack an agent loads at session start contains only must and should entries — context entries are deliberately excluded to keep the pack small and rule-focused. Within each scope, MUST entries are ordered ahead of SHOULD, and when the ~1,500-token cap is reached the lower-priority tail is dropped first. The result: the binding rules are what survive, always. The pack then closes with a ## Tags line — the distinct tags across the lore in scope — as a map of which topics have governed knowledge to search before acting; it is budgeted inside the same cap and truncated last, so it never displaces a rule.
context entries are still fully searchable — an agent finds them with lore_search when a question calls for them. They just don't consume the standing pack budget.
How priority affects search
Search ranks by text relevance plus a priority boost, so among comparably relevant matches a MUST rises above a SHOULD, which rises above context. Priority is a ranking signal, not a filter — a strongly matching context entry still surfaces.
What agents do with priority
The agent skill instructs agents to treat MUST entries as non-negotiable and to cite the ids they relied on, so a human can trace any decision back to the governing rule. SHOULD entries are followed by default; deviating from one is allowed but should be deliberate and explained. Choosing the right priority when you author an entry is therefore how you set the strength of the instruction — reserve must for the genuine non-negotiables.