Content
71%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.
This is an exemplar progressive-disclosure skill — a concise router that keeps every domain detail in 25 real, well-signaled reference files loaded one level deep. Its weaknesses are repetition of the activation mandate (stated three times in the body after already appearing in the description), over-long table cells that leak reference-file parameter detail into the overview, and workflow validation steps that name checks without defining how to run them.
Suggestions
State the load-mandate once: the description already carries 'MANDATORY... do not answer from training knowledge', so collapse the three repeated directives in the Reference Files section into a single line.
Trim config-key specifics out of 'When to Load' cells (e.g. READER_EMPTY_RECORDS_FETCH_INTERVAL, SHARD_GET_RECORDS_MAX, JOB_MANAGED vs SELF_MANAGED) — keep goal-level conditions in the table and leave parameter names inside the reference files.
Make workflow validation steps concrete: replace 'Validate against best practices' and 'Compile and test locally' with the specific check or command (e.g. the compile command from environment-setup.md) and add a failure branch.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly a lean routing table, but the activation mandate is repeated three times ('You MUST use this skill and its reference files...', 'Do NOT answer from training knowledge...', 'You MUST load the relevant reference files below') after already appearing in the description, and several 'When to Load' cells embed config-key-level detail (e.g. 'READER_EMPTY_RECORDS_FETCH_INTERVAL', 'SHARD_GET_RECORDS_MAX', 'JOB_MANAGED vs SELF_MANAGED') that belongs inside the reference files. Mostly efficient but could be tightened — anchor 3, not the minor-trims-only anchor 4. | 3 / 5 |
Actionability | The two numbered workflows give concrete routing instructions with exact READ steps and file links, and the table specifies precisely when to load each of 25 files; for an instruction-only steering skill this is mostly executable guidance. Minor gaps: 'Validate against best practices' and 'Compile and test locally' name steps without giving a method or command, keeping it below the copy-paste-ready anchor 5. | 4 / 5 |
Workflow Clarity | Both workflows are clearly sequenced with distinct paths for new applications vs. general questions, and validation steps exist ('Validate against best practices', 'Compile and test locally'), so the sequence is well above anchor 3's implicit-checkpoint level. It falls short of anchor 5 because the checkpoints are named but not explicit — no failure branch or feedback loop (what to do when validation or compilation fails) is described. | 4 / 5 |
Progressive Disclosure | The body is a pure overview/router: roughly 85 lines with zero inlined domain content, a complete goal-to-reference table covering all 25 reference files (all of which exist on disk), one-level-deep references with 'When to Load' conditions, and negative-scope notes guiding navigation (the Iceberg rows explicitly redirect catalog/maintenance questions to iceberg-tuning-and-operations.md). Cross-links inside reference files are sibling pointers, not nested content, so navigation is easy and unambiguous. | 5 / 5 |
Total | 16 / 20 Passed |