Conciseness

The ~90-line body is lean and every section carries distinct information: problem framing, a compact tier table, gate thresholds, runnable commands, hook config, and hard rules. It assumes Claude's competence (never explains what CLAUDE.md or hooks are) and includes no padding, matching the "every token earns its place" anchor. The 6-line "Forcing questions" section is the only arguably trimmable material, but it is brief and purposeful.

Actionability

Guidance is concrete: exact commands with flags ("python3 scripts/memory_inspect.py --why \"PR base branch is dev\"", "--contested"), a dry-run promotion pass, named hooks, and per-hook env-var kill switches ("AGENT_MEMORY_SESSIONEND=0"). It falls short of fully executable because the referenced scripts (scripts/memory_inspect.py, scripts/memory_promote.py) and the /cs:memory adopt command are not present in the bundle, so the commands cannot actually run as shipped.

Workflow Clarity

The promotion lifecycle is clearly sequenced by tier (L0→L1→L2→L3) with explicit gate criteria ("≥ 3 distinct sessions spanning ≥ 2 distinct calendar days"), refusal conditions, and safety checkpoints ("dry-run the promotion pass; writes nothing", adopt "backs both up first", hooks "fail open"). The gap keeping it below a 5 is that the operational loop is assembled implicitly across sections — the human's adopt step is referenced ("/cs:memory adopt") but never shown as a runnable step.

Progressive Disclosure

Section structure is good and the single external pointer is clearly signaled with its contents named ("Rationale, open decisions, field schema: DESIGN.md"), keeping the body appropriately short. However, the bundle contains no scripts/, references/, or DESIGN.md — the referenced script paths and the ../../DESIGN.md pointer (which also escapes the skill directory) resolve to nothing, so navigation is broken as shipped rather than merely imperfectly organized.

Specificity

The description lists several concrete capabilities — "Implements a four-tier store (L0 transcripts / L1 candidates / L2 project context / L3 stable persona)", promotion "earned by recurrence across sessions and days", and "nothing reaches a committed file without a human adopting it" — which would merit a 4. However, the trigger clause is framed in second person ("you want the agent to learn durable facts"), which the judging guidelines penalize by reducing specificity by 1.

Completeness

Explicitly answers both questions: the "what" is stated concretely ("Implements a four-tier store... promotion is earned by recurrence... nothing reaches a committed file without a human adopting it") and the "when" opens with "Use when a project's CLAUDE.md has grown past what anyone reads" plus three concrete alternative trigger phrases. This matches the anchor-5 example pattern of "concrete what + explicit 'Use when' with concrete trigger phrases".

Trigger Term Quality

Good coverage of natural phrases a user would actually say: "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". A few natural synonyms are missing (e.g., "remember this", "forget", "persistent context"), which keeps it below the comprehensive anchor 5.

Distinctiveness Conflict Risk

The niche is clear and the triggers are distinctive — session-derived memory with recurrence-gated promotion and provenance questions like "where a memory line came from" are unlikely to fire for an unrelated skill. Minor overlap risk remains with other memory/CLAUDE.md-management skills, since the description does not itself differentiate from e.g. a manually-written knowledge vault.