Content
78%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.
The SKILL.md body is a well-organized overview that leans on decision tables, executable CLI commands, and one-level-deep reference files, with strong progressive disclosure and good actionability. The main improvement lever is tightening the repeated maxTokens guidance across sections.
Suggestions
Consolidate the maxTokens/quota warning: state it once in Critical Warnings and reference it from the Converse, Troubleshooting, and quota examples instead of restating it verbatim in each.
Inline at least one complete validate-then-proceed feedback loop (e.g., for the Harness create→poll→invoke flow) so the body demonstrates the validation pattern, not just names the steps.
Move the 'Examples — mapping user intent to workflows' list toward the decision-table format used in 'Which Bedrock Capability Do You Need?' to reduce prose tokens and improve scannability.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely lean and assumes Claude's competence — terse tables, runnable commands, and pointed warnings — but the maxTokens caution is repeated across Critical Warnings, Converse API, Troubleshooting, and multiple examples, adding some redundancy that could be trimmed without losing the safety signal. | 4 / 5 |
Actionability | Concrete, copy-paste-ready AWS CLI commands appear throughout ('aws bedrock-runtime converse --model-id us.anthropic.claude-sonnet-4-6 ...', 'aws service-quotas get-service-quota ...') with a real retry-config snippet, and detailed code is correctly routed to reference files; minor gaps exist because most full procedures live in references rather than inline. | 4 / 5 |
Workflow Clarity | Multi-step processes are explicitly sequenced and routed to reference procedures with 'execute step by step' / MUST constraints, and key checkpoints are named (poll get-harness until READY, run prepare-agent after changes, confirm before destructive ops); full validation feedback loops mostly live in the referenced procedures rather than inline, keeping this just below a 5. | 4 / 5 |
Progressive Disclosure | A clear overview with a TOC and goal→reference decision tables routes each task to a single one-level-deep reference file, all 22 referenced files verified to exist, with detail appropriately split out and easy to navigate. | 5 / 5 |
Total | 17 / 20 Passed |