Content
86%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 body is an exemplar of lean, fully executable CLI reference material with a clear core workflow and useful decision guidance on when to prefer it over the built-in browser tool. Its main structural weakness is that the entire command reference lives inline in SKILL.md rather than being split into a one-level-deep reference file, which would trim the always-loaded token cost.
Suggestions
Move the exhaustive per-category command listing (Navigation through Tabs & Frames) into a references/COMMANDS.md and keep only the core workflow, key commands, and examples in SKILL.md.
Add a brief error-recovery note (e.g., re-snapshot and re-select refs when a ref becomes stale after page changes) to close the workflow feedback-loop gap.
State where the snapshot refs come from in the Core Workflow comments is already good; consider one line on interpreting the refs JSON for less common element roles.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is almost entirely dense, commented command blocks and tables of commands — it explains nothing Claude already knows and contains no padding. It matches anchor 5 ('lean and efficient; assumes Claude's competence') rather than anchor 4, which reserves room for over-explanation that is simply absent here. | 5 / 5 |
Actionability | Every section is copy-paste-ready CLI syntax ('agent-browser click @e2', 'agent-browser fill @e3 "text"'), and two worked examples (search-and-extract, multi-session testing) cover common cases. This matches anchor 5's 'fully executable; copy-paste ready... specific examples cover the common cases'. | 5 / 5 |
Workflow Clarity | The Core Workflow section sequences navigate → snapshot → parse refs → interact → re-snapshot, reinforced by stability checkpoints ('wait --load networkidle') in Best Practices and the examples. It falls short of anchor 5 because there are no explicit validate→fix→retry feedback loops (e.g., what to do when a ref is stale or a wait times out), though no destructive/batch operation triggers the score-3 cap. | 4 / 5 |
Progressive Disclosure | The body is well-sectioned but is a ~200-line full command API reference inlined entirely in SKILL.md, with no references/, scripts/, or assets/ bundle to offload it. This matches anchor 3 ('some structure... content that should be separate is inline') — better organized than anchor 2's structureless wall, but lacking anchor 4's 'bulk in separate file' split. | 3 / 5 |
Total | 17 / 20 Passed |