Conciseness

The body is dense with material Claude does not already know — the tier/injection table, numeric gates ("≥ 3 distinct sessions spanning ≥ 2 distinct calendar days"), env-var toggles — and explains no basic concepts. Two sections could be trimmed: the "The problem" narrative and especially "Forcing questions", which addresses the human reader rather than instructing the agent. That is more than zero padding but well short of the anchor-3 "some unnecessary explanation" level.

Actionability

Every user-facing command is copy-paste ready with real flags: "python3 scripts/memory_inspect.py --tier L1", "--why \"PR base branch is dev\"", "--contested", and "python3 scripts/memory_promote.py" marked "dry-run the promotion pass; writes nothing"; both referenced scripts exist in the bundle. Hook control via AGENT_MEMORY_SESSIONSTART=0-style env vars and the explicit "/cs:memory adopt" command cover the remaining cases.

Workflow Clarity

The L0→L3 ladder is a clear sequence with explicit gates, and validation checkpoints are present for the destructive step (dry-run promotion, staged proposals in ".memory/staged/", "it backs both up first" on adopt, quarantine of redacted claims, freeze on contradiction) — so the destructive-operation cap does not apply. What keeps it below 5 is that no single end-to-end walkthrough (capture → inspect → promote → adopt) ties the sections together, and hook setup is only described, not shown.

Progressive Disclosure

The body itself is well-sectioned and appropriately lean, but the reference layer is disconnected: the only detail link, "Rationale, open decisions, field schema: ../../DESIGN.md", points outside the bundle rather than to a bundled file, while the three files in references/ (hook_capture_discipline.md, promotion_gate_design.md, tiered_memory_canon.md) and assets/memory_schema.json are never referenced from SKILL.md at all. Deep material exists but is undiscoverable from the entry point — structure is present, signaling is not.

Specificity

Concrete mechanisms are named — "Implements a four-tier store (L0 transcripts / L1 candidates / L2 project context / L3 stable persona) where promotion is earned by recurrence across sessions and days" and "nothing reaches a committed file without a human adopting it" — which alone would merit a 4. However, the trigger clause "you want the agent to learn durable facts from its own sessions" is second-person voice, which the rubric penalizes by one point on this dimension. Coverage also omits the concrete operations the skill offers (inspect, provenance queries, adopt).

Completeness

Both halves are explicit: what — "Implements a four-tier store (L0 transcripts / L1 candidates / L2 project context / L3 stable persona) where promotion is earned by recurrence... and nothing reaches a committed file without a human adopting it"; when — an explicit "Use when..." clause with three concrete trigger scenarios. This matches the anchor-5 example's structure.

Trigger Term Quality

The triggers read as natural user utterances: "a project's CLAUDE.md has grown past what anyone reads", "why the agent keeps re-learning the same correction", "why a remembered rule is wrong", "where a memory line came from". They include the artifact name (CLAUDE.md) and phrasings a user would actually say across the skill's scenario space, with no jargon-only gaps.

Distinctiveness Conflict Risk

The niche is clear and the symptom-based triggers (CLAUDE.md bloat, memory-line provenance, re-learning corrections) are distinct, but the first trigger clause — wanting "the agent to learn durable facts from its own sessions" — is broad enough to also fire for closely related session-memory or memory-audit skills. Minor overlap risk, matching the anchor-4 example.