Content
62%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
This is a solid instructional skill with excellent workflow clarity — the core loop pattern is well-defined and consistently applied, with good error recovery guidance. However, it's somewhat verbose for its audience (explaining concepts like pipes, shell state, and control characters that Claude already knows) and uses a pseudo-notation for examples rather than actual executable tool calls, which limits actionability. The content would benefit from trimming known concepts and either using real tool call syntax or explicitly noting the abstraction layer.
Suggestions
Trim sections explaining concepts Claude already knows (control characters, what pipes are, that shell state persists, that rm is destructive) to brief reminders or remove entirely — this could cut 30-40% of tokens.
Use actual wsh tool call syntax (e.g., `wsh_send_input({"input": "npm install\n"})`) in at least one complete end-to-end example rather than the abstract `send input:` notation throughout.
Consider splitting 'Common Patterns' and 'Pitfalls' into a separate PATTERNS.md reference file, keeping SKILL.md focused on the core loop and essential guidance.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The content is generally well-written but includes some unnecessary elaboration that Claude would already know — e.g., explaining what pipes are, that shell state persists, what control characters do, and that `rm` does real damage. Several sections could be tightened significantly while preserving all useful information. | 2 / 3 |
Actionability | The skill provides concrete patterns with inline examples (e.g., `send input: npm install\n`, heredoc patterns, pager bypasses), but the examples use a pseudo-notation (`send input:`, `wait for idle`, `read screen`) rather than actual executable tool calls or API commands. The execution context preamble tries to bridge this gap but the body itself remains in an abstract instruction language. | 2 / 3 |
Workflow Clarity | The core loop (send → wait → read → decide) is clearly articulated and consistently reinforced throughout. The chained commands section explicitly recommends separate cycles for error detection, the long-running commands section describes a polling feedback loop, and the destructive commands section includes validation checkpoints (confirm, double-check, dry-run). Error recovery strategies are well-sequenced (Ctrl+C → Ctrl+Z → kill → ask human). | 3 / 3 |
Progressive Disclosure | The content is well-organized with clear section headers and logical grouping, but it's a long monolithic document (~200+ lines) with no references to external files. The execution context preamble references `skills/core/SKILL.md` for HTTP fallback, but the main body could benefit from splitting detailed patterns (e.g., Common Patterns, Pitfalls) into separate reference files. | 2 / 3 |
Total | 9 / 12 Passed |