Content
81%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.
A dense, operational skill body with excellent error handling, explicit expected outcomes, and strong safety rules. Its weaknesses are moderate: a verbatim-duplicated CDP startup block, and core click/type syntax being deferred entirely to a runtime-fetched upstream guide, which costs it both actionability and conciseness.
Suggestions
Consolidate the duplicated CDP launch/polling script: define the 'ensure CDP is up' block once (e.g., a single copy in the CDP Startup Contract section) and reference it from the Smoke Path instead of repeating ~20 lines verbatim.
Inline the handful of core interaction commands (click, type, wait, element screenshot syntax) in a short reference table so the most common path does not require shelling out to `agent-browser skills get core` before every interaction.
Replace the hedged 'when the core guide exposes an element-screenshot command' with the actual command name and arguments once verified, making element screenshots copy-paste ready.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and command-first — it assumes competence, explains no general browser concepts, and every section earns its tokens. It is not a 5 because the full CDP launch/polling block (~20 lines) is duplicated verbatim in 'CDP Startup Contract' and 'OpenDesign Smoke Path', and the Chrome manual-launch command also appears twice, which could be consolidated. | 4 / 5 |
Actionability | Most guidance is copy-paste ready: install check, connect sequence with curl polling, snapshot/screenshot commands, cleanup trap, and expected success criteria. It is not a 5 because core interaction commands (click, type, wait) are never shown — they are deferred entirely to `agent-browser skills get core` — and the element screenshot is hedged as 'when the core guide exposes an element-screenshot command', leaving the common click/type path without concrete syntax inline. | 4 / 5 |
Workflow Clarity | The 11-step Workflow is clearly sequenced with explicit checkpoints: verify install, confirm CDP before connecting, snapshot before selecting elements, re-snapshot after state changes, and report title/URL/text/screenshot. Error recovery is specified for each failure mode (Chrome crash message, manual-launch fallback, stop if CLI missing), and the smoke path defines expected success outputs (title 'OpenDesign', URL under 127.0.0.1:17573, screenshot path). This matches the anchor-5 feedback-loop pattern; a 4 would lack the fallback and expected-result definitions. | 5 / 5 |
Progressive Disclosure | Structure is good: well-labeled sections, and heavy upstream material (electron, slack, dogfood, sandbox, agentcore guides) is loaded on demand via `agent-browser skills get` and redirected to temp files — a genuine one-level-deep disclosure pattern. It is not a 5 because the core command reference is only reachable through a runtime CLI call rather than a clearly signaled in-bundle reference, and the duplicated CDP block suggests content that belongs in one place, leaving minor organization gaps. | 4 / 5 |
Total | 17 / 20 Passed |