Conciseness

The operational core is lean (commands, flag table, config keys), but the conceptual intro ("the deployment-time analogue of training: short-term experience → long-term competence") and the "synthesizes three ideas" section (SkillOpt / Claude Dreams / Agent sleep) explain background Claude does not need to run the skill. This is minor, localized over-explanation — the 4 anchor — rather than the pervasive padding of 3.

Actionability

Commands are copy-paste ready across all operations: "${CLAUDE_PLUGIN_ROOT}/scripts/sleep.sh" status/dry-run/run/adopt, schedule/unschedule, plus a full 14-row flag table with defaults, config keys, and a deterministic validation command (python -m skillopt_sleep.experiments.run_experiment --persona researcher --assert-improves). Matches the 5 anchor for fully executable coverage of common cases.

Workflow Clarity

The six-stage cycle (Harvest -> Mine -> Replay -> Consolidate -> Stage -> Adopt) is clearly sequenced, and validation around the destructive adopt step is explicit: held-out gate, staging where "Nothing live changes", backups on adopt, and "Always show the user the held-out baseline → candidate score ... Evidence before adoption." This matches the 5 anchor with explicit validation and error-recovery guidance (dry-run).

Progressive Disclosure

Sections are well organized and the skill is effectively self-contained (no references/ bundle exists; external detail is one-level-deep via the upstream guide URL). However, the full CLI flag table and config-key reference are inlined where a references/ file would serve better, and the closing meta-note about the non-vendored design-doc path is README material. Fits the 4 anchor: good structure, minor organization gaps.

Specificity

The description lists a concrete action pipeline — "harvest past sessions -> mine recurring tasks -> replay offline -> consolidate validated CLAUDE.md and SKILL.md behind a held-out gate" — giving comprehensive, specific coverage of what the engine does, in third-person voice. It matches the 5 anchor rather than 4 because the what is covered completely, with no notable gaps.

Completeness

Both halves are explicit: "Use when the user wants their Claude agent to self-improve... or says things like..." answers when with concrete trigger phrases, and "Drives the skillopt_sleep engine: harvest -> mine -> replay -> consolidate" answers what. This is a direct match for the 5 anchor.

Trigger Term Quality

It includes a wide range of natural user phrasings: "make my agent better the more I use it", "review my past sessions", "learn my preferences", "consolidate what you learned", "run the sleep cycle", plus synonyms like nightly/offline "sleep" or "dream" cycle. Coverage is comprehensive with variants, matching the 5 anchor.

Distinctiveness Conflict Risk

The sleep/dream-cycle framing and skillopt_sleep naming are a clear niche, but phrases like "learn my preferences", "memory/skill consolidation", and "consolidate ... CLAUDE.md" carry minor overlap risk with generic memory-management skills. Fits the 4 anchor (mostly distinct, minor overlap with closely related skills) better than 5.