Conciseness

The body is dense and assumes competence (no padding about what parquet is or how libraries work), but the Gotchas section repeats body material nearly verbatim: the static-column error appears twice, and the '/__properties' drop, prefix-stripping rule, and 'no pandas munging' advice each appear in both a main section and a gotcha. This fits the 'efficient; minor instances that could be trimmed' anchor rather than the every-token-earns-its-place anchor.

Actionability

Guidance is fully executable: a complete stream() call with every keyword argument, a builder table with argument orders, and two copy-paste-ready examples with imports (end-to-end Transform3D, Points3D via to_packed_component) covering the common cases. This matches the top anchor.

Workflow Clarity

The sequence is present and stated ('fill in the rerun-data-model mapping table first, then express it through stream()'), and validation behavior is explicit (config validated against the file's schema at stream(), ValueError raised eagerly for missing columns). It falls short of the top anchor because there is no explicit ordered workflow or validate-fix-retry loop; it stays above score 3 because checkpoints (eager validation, lazy static-column error as data-quality signal) are described rather than absent.

Progressive Disclosure

There are no bundle files, but the body is well-sectioned and the References section gives clearly signaled, one-level-deep pointers (SDK source paths, integration test files, sibling skills) with no nested references. It matches 'good structure; most content appropriately placed' — the lens builder table and selector grammar could arguably live in a separate reference file given the ~245-line body, keeping it below the top anchor, but nothing is buried or inline that clearly belongs elsewhere.

Specificity

The description lists multiple concrete actions — 'Ingest tabular Parquet files into Rerun chunk streams', 'column grouping, timeline/index columns, static columns', and 'lenses (DeriveLens) that assemble the typed components (Transform3D, Scalars)' — with comprehensive coverage of the skill's capabilities, matching the anchor for multiple specific concrete actions. It exceeds the score-4 anchor because coverage is comprehensive rather than having minor gaps.

Completeness

It explicitly answers both questions: what ('Ingest tabular Parquet files into Rerun chunk streams with rerun.experimental.ParquetReader') and when ('Read when converting trajectory or sensor tables (LeRobot-style parquet, exported logs) into entities and components') with concrete trigger phrases. Score 4 would require the 'when' to be less explicit, which it is not.

Trigger Term Quality

Natural user vocabulary is covered with synonyms for both format and data type: 'Parquet files', 'parquet', 'tabular', 'trajectory or sensor tables', 'LeRobot-style parquet', 'exported logs'. This matches the comprehensive-coverage anchor; the only missing variation is the '.parquet' file extension, which the anchor-4 example's thinner coverage does not fit better.

Distinctiveness Conflict Risk

The parquet-to-Rerun ingestion niche is clear, but the description shares trigger vocabulary ('rerun', 'chunks', 'components', 'lenses') with the closely related skills it names (rerun-chunk-processing, rerun-data-model), creating minor overlap risk. It is not score 5 because those sibling skills could plausibly compete for the same Rerun-centric queries, though the role division is delineated.