Conciseness

Mostly efficient but with noticeable tightening opportunities: the kg_supersede/kg_invalidate/kg_add selection guidance appears verbatim in both "Protocol" step 6 and "Unhappy paths"; "wrong is worse than slow" and "do not silently fall back to model memory" each appear twice; and the "Identity" persona section ("Act as a senior AI-memory systems engineer with decades of experience...") is padding Claude does not need. This fits 'includes some unnecessary explanation or could be tightened' rather than the 'minor instances' of a 4.

Actionability

Concrete, executable guidance throughout: named MCP tools (mempalace_search, mempalace_kg_query, mempalace_diary_read...), a tool-selection table, specific commands (mempalace --version, mempalace status), and concrete query-shape guidance ("short natural-language query (keywords or a question — not a system prompt)"). Falls short of a 5 because there is no worked example of an actual query, and detail is deferred to external URLs the agent may not fetch.

Workflow Clarity

Clear sequenced protocol (Step 0 availability check → injected memory → pre-response search → verify when unsure → return verbatim → diary write → supersede on change) with explicit validation up front and a dedicated "Unhappy paths" section providing error-recovery feedback loops (empty results → widen scope per the query guide; MCP errors → surface and suggest mempalace status). This matches the top anchor's explicit validation plus feedback loops.

Progressive Disclosure

No bundle files exist (references/, scripts/, assets/ absent), and the body is well sectioned with clearly signaled, one-level-deep external pointers (the canonical recall-protocol URL and query guide). Minor gap: the body runs ~120 lines with the full protocol, tool table, and anti-patterns all inline — some detail could be split out — keeping it below the 'clear overview with appropriately split content' anchor.

Specificity

The description names the domain ("Recall protocol for MemPalace") and one concrete action ("search the palace before answering about past work, people, projects, or prior decisions"), matching the anchor for 1-2 concrete actions without comprehensive coverage. It does not list several specific actions (the individual tools like search/kg_query/diary_read stay in the body), so it is not a 4.

Completeness

It explicitly answers both: what ("search the palace before answering about past work, people, projects, or prior decisions") and when ("Apply when the user asks what was decided, what happened before, who someone is, what was discussed last time..."). The explicit 'Apply when' clause with concrete trigger phrases matches the top anchor exactly; a 4 would require a weaker or less specific 'when', which is not the case.

Trigger Term Quality

It comprehensively covers natural user phrasings — "what was decided", "what happened before", "who someone is", "what was discussed last time", "past work, people, projects, or prior decisions", "anything that may already be filed in their memory palace" — plus the explicit invocation name. These are the exact forms a user would naturally say, matching the comprehensive-synonyms anchor; no neighboring anchor fits better.

Distinctiveness Conflict Risk

It carves a clear niche (recall only) and explicitly disambiguates from the sibling skill ("Complements the mempalace setup skill and requires the mempalace-mcp server"), with the body stating the mempalace skill covers install/mine/status while this one covers recall. Distinct triggers and explicit boundary mean minimal conflict risk.