Content
85%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
An expert-grade ingestion skill: a clear two-path decision (view vs convert), sequenced steps with real validation checkpoints for batch downloads and FK mapping, and bundle scripts that absorb the fragile logic. The only weaknesses are minor — a couple of narrative asides and two code snippets (zero_pad, ParquetReader) that are not quite copy-paste executable.
Suggestions
Make the splitting snippet fully executable by replacing the pseudo-helper zero_pad(entry.recording_id) with a concrete expression (e.g. f"episode_{int(entry.recording_id.split('_')[1]):05d}") or by defining the helper inline.
Complete the ParquetReader example with its import and the undefined paths (episode_parquet, rrd_dir) so the FK source snippet is copy-paste ready, matching the quality of the LeRobotReader and snapshot_download examples.
Trim the few motivational asides (e.g. 'which is usually not what the user pictured when they asked to look at the data', 'Do the equivalent by hand and three things bite' framing) to tighten the body without losing the operational content.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and load-bearing — it explains only non-obvious facts (v2.1 vs v3.0 file layout, zero-padding rationale, send_properties=False, why the fetch script asserts) and assumes Claude's competence rather than re-explaining basics. A few narrative asides ("usually not what the user pictured", "three things bite") could be trimmed, which matches 'efficient; minor instances of over-explanation' rather than the every-token-earns-its-place level 5. | 4 / 5 |
Actionability | Concrete runnable commands (uv run scripts/fetch_lerobot_episodes.py <repo_id> --episodes 5, the HF_TOKEN check) and near-complete Python snippets cover the common paths, but minor gaps remain: zero_pad(entry.recording_id) is a pseudo-helper rather than a real API, and ParquetReader's import plus rrd_dir/combined_rrd are undefined in their snippets. This sits between 'minor gaps' (4) and 'copy-paste ready' (5), closer to 4 because two snippets are not directly executable as written. | 4 / 5 |
Workflow Clarity | Steps are explicitly sequenced (Step 0 pick/size/token check → Step 1 combined RRD → Step 2 per-episode → Step 3 URDF) with a viewing-vs-converting decision up front, and validation is present for the batch operations: the fetch script "verifies that exactly the requested episodes landed" and asserts, sizes are reported before download, and check_joint_mapping.py gates trusting an FK pose, with a Gotchas checklist covering failure modes and recovery. Not 4, because checkpoints and error-recovery guidance are explicit rather than merely present. | 5 / 5 |
Progressive Disclosure | The overview body keeps decision logic and minimal code while the heavy lifting is split into two real, well-signaled bundle scripts (both present in scripts/ and described with their behavior and output), and remaining references (rerun-chunk-processing, rerun-urdf, the external dataloader example) are one level deep with clear scope hints. Not 4: there are no buried references and no content that clearly belongs in a separate file inlined in the body. | 5 / 5 |
Total | 18 / 20 Passed |