Content
53%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.
Highly actionable content with executable commands and some well-structured workflows, including a real error-recovery loop. It is undermined by severe verbosity from duplicated examples and scattered version-dated material, a monolithic structure with zero progressive disclosure into bundle files, and batch destructive workflows that lack validation checkpoints.
Suggestions
Split reference material into bundle files — e.g. references/models.md (the model catalog with pricing), references/config.md (config.toml profiles and environment variables), and references/ci-cd.md — keeping SKILL.md as a concise overview with clearly signaled one-level-deep links, and remove the 'Related Skills' section or point it at files that actually exist.
Deduplicate examples: state the bypass/full-auto/sandbox flags once in an options section and reference them thereafter instead of repeating '--dangerously-bypass-approvals-and-sandbox' in ~20 examples; consolidate the three overlapping model-selection sections into one.
Consolidate time-sensitive details (GPT-5.2 release date, pricing, knowledge cutoffs) into a single clearly-labeled versions section, and add explicit validation checkpoints (e.g. run tests and review diffs before committing) to the batch destructive workflows that currently run unattended under bypass mode.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The same flags and commands repeat across sections ('--dangerously-bypass-approvals-and-sandbox' appears verbatim in roughly 20 examples) and model-selection guidance is duplicated across three sections. Time-sensitive material ('December 2025', GPT-5.2 pricing, knowledge cutoff dates) is scattered throughout instead of quarantined in a versions/deprecated section, which the rubric explicitly penalizes. It is command-dense rather than re-teaching known concepts, keeping it above anchor 1 but firmly at anchor 2. | 2 / 5 |
Actionability | Nearly everything is copy-paste executable: install/auth commands, full bash workflow functions (e.g. 'refactor_module' with git stash backup and restore on failure), a complete GitHub Actions YAML, and config.toml profile examples. Specific examples cover the common cases, matching anchor 5. | 5 / 5 |
Workflow Clarity | Sequences are clearly laid out (Think-Act-Observe loop, 6-step task decomposition) and the refactoring workflow has a genuine feedback loop ('if [ $? -eq 0 ] ... else git stash pop'). However, many destructive/batch operations ('Convert all .js files to TypeScript', 'Run tests and auto-fix all failures', 'Achieve 100% coverage') run under bypass mode with no validation checkpoint, which caps this dimension at 3 per the rubric's batch-operations guideline. | 3 / 5 |
Progressive Disclosure | The body is a 687-line monolith with no bundle files at all (no references/, scripts/, or assets/ exist), and material that clearly belongs in separate files — a ~100-line model catalog, ~80-line config reference, CI/CD pipeline example, and troubleshooting — is fully inlined. The 'Related Skills' section points to skills ('codex-auth', 'codex-git', etc.) that are not present. Section headers provide some structure, but the scale of inlining matches anchor 2 better than anchor 3. | 2 / 5 |
Total | 12 / 20 Passed |