Conciseness

The body is dense and earns its tokens — the tier table, the gates, and the hard rules carry information Claude cannot infer, with commands shown compactly under "Use it". It is not a 5 because the "The problem" narrative ("The fix is not more storage — it is a promotion ladder") and the four "Forcing questions" are rhetorical framing that could be trimmed without losing operational content.

Actionability

Concrete, executable guidance dominates: `python3 scripts/memory_inspect.py --tier L1`, `--why "PR base branch is dev"`, `--contested`, the dry-run `python3 scripts/memory_promote.py`, and exact env-var toggles (`AGENT_MEMORY_SESSIONSTART=0` etc.). Not a 5 because the pivotal `/cs-memory adopt` command is referenced ("Only an explicit `/cs-memory adopt` touches a `CLAUDE.md`") but never shown or defined, leaving the final, most consequential step without an executable form.

Workflow Clarity

The lifecycle is well-specified with strong safety checkpoints: staged promotions, dry-run that "writes nothing", `--contested` surfacing frozen claims, backup before adopt ("it backs both up first"), and fail-open hooks. Not a 5 because the operator-facing sequence (capture → inspect → resolve contested → promote → adopt) must be assembled from the tier table, gates, and "Use it" sections rather than being laid out as one ordered walkthrough with error-recovery branches.

Progressive Disclosure

The "More in this skill" section clearly signals all three reference files (which exist in the bundle) with one-line descriptions at one level of depth, plus script entry points — good structure. Not a 5 because the rationale pointer `[../../DESIGN.md](../../DESIGN.md)` escapes the skill bundle two levels up and does not resolve, and `assets/memory_schema.json` is only reachable indirectly via that link and a validate script mention rather than being directly signposted.

Specificity

The description names the domain and concrete mechanisms — "Promotes facts through four tiers only on recurrence across sessions and days, with memory_promote.py and a human adopt step" — which sits at the 4 anchor (several specific actions, minor gaps), but the second-person phrase "or you need to trace where a remembered rule came from" triggers the rubric's explicit one-point voice penalty, dropping it to 3. Not a 2 because the actions named are concrete and domain-specific, not generic.

Completeness

Both halves are explicit: the "what" is "Promotes facts through four tiers only on recurrence across sessions and days, with memory_promote.py and a human adopt step", and the "when" opens with "Use when an agent keeps relearning the same correction, CLAUDE.md has grown unreadable, or you need to trace where a remembered rule came from" — three concrete trigger phrases. This matches the 5 anchor exactly; a 4 would require the when-clause to be less explicit or specific, which it is not.

Trigger Term Quality

Natural user phrasings are present — "keeps relearning the same correction", "CLAUDE.md has grown unreadable", "trace where a remembered rule came from" — matching the 4 anchor of good keyword coverage with a few natural terms missing. Not 5 because common synonyms like "agent keeps forgetting instructions" or "memory system" are absent, and no file/extension-style triggers appear.

Distinctiveness Conflict Risk

The description carves a clear niche — tiered promotion of agent memory with deterministic recurrence gates — and its triggers ("relearning the same correction", "CLAUDE.md has grown unreadable") would not fire for unrelated skills. Minimal conflict risk, matching the 5 anchor; not lower because nothing here overlaps generically with other skill domains.