Conciseness

Every token carries package-specific facts Claude cannot know ("RUVECTOR_RABITQ_THRESHOLD memories (default 5000)", "about 0.5 to 3 ms instead of a full scan (about 9 ms at 5k, 38 ms at 20k)", "rebuilds once those changes reach about 10% of its size") with zero padding or explanation of general concepts; even the tangential closing install section is only 7 lines, matching the lean/every-token-earns-its-place anchor.

Actionability

Guidance is fully executable and copy-paste ready throughout: exact CLI invocations with flags ("npx ruvector diskann build ./idx -i vectors.json [--max-degree 64] [--alpha 1.2] [--json]"), the input file format ("a JSON array of { \"id\": \"...\", \"vector\": [ ... ] }"), named MCP tools with parameter lists, and concrete library entry points ("buildDiskAnnIndex(dir, records)", "searchDiskAnnIndex(dir, query, k)").

Workflow Clarity

Sequences are clear — the RaBitQ recall flow is literally a numbered two-step process ("1. A 1-bit RaBitQ index proposes a candidate pool... 2. The pool is re-scored..."), DiskANN's build-then-search order is encoded in the commands, and verification affordances exist ("getStats().recallEngine reports 'rabitq' or 'brute-force'", the `engine` output field, `--json` flags). It falls short of the top anchor because there is no explicit validate-and-recover guidance for the build/index operations (e.g. what to do when a diskann build rejects malformed input), keeping it at clear-sequence/most-checkpoints rather than explicit-feedback-loops.

Progressive Disclosure

The body (~100 lines) is well organized into per-accelerator sections with a clean extras table and no bundle files exist to reference, so nothing is misfiled — but the detailed behavioral specs (e.g. the ~20-line RaBitQ build-timing/rebuild rules) sit inline with no one-level-deep pointers to deeper material, fitting the good-structure/minor-gaps anchor rather than the clear-overview-with-signaled-references top anchor.

Specificity

The description enumerates the full set of concrete capabilities — "RaBitQ-accelerated hook memory recall, the DiskANN on-disk index (CLI and MCP), MinCut/Louvain/spectral code-graph analysis, and `npx ruvector add` extras" — matching the multiple-specific-actions, comprehensive-coverage anchor rather than the minor-gaps anchor at 4.

Completeness

It explicitly answers both what ("Use the optional native and WASM accelerators that the `ruvector` npm package wires in...") and when ("Use when a task involves large hook memory stores...") with concrete trigger phrases, plus an explicit negative boundary ("Do not use for generic vector-database questions"), matching the top anchor.

Trigger Term Quality

Natural phrases like "large hook memory stores", "building or searching a vector index on disk", "finding module boundaries or communities in a codebase", and "deciding which ruvector extra to install" give good coverage, but common synonyms and variations (e.g. "vector search", "code graph", "recall speed") are missing, which fits the good-coverage/few-missing anchor at 4 rather than the comprehensive-synonyms anchor at 5.

Distinctiveness Conflict Risk

The skill is pinned to a niche package (`ruvector` extras/accelerators) with distinct triggers, and the "Do not use for generic vector-database questions that the core `ruvector` API already answers" clause explicitly separates it from the core-API skill, giving minimal conflict risk.