cua-driver

Drive a native GUI app (macOS, Windows, Linux) via the Qwen Cua Driver CLI (default) or MCP server; snapshot its accessibility tree, act through snapshot-bound element tokens, native menu paths, exact window geometry, or pixel coordinates, and verify from fresh state. Use when the user asks you to operate, drive, automate, or perform a GUI task in a real application on the host.

Quality

85%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

77%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

Highly actionable and workflow-clear, with executable examples and a well-sequenced verify-then-escalate ladder. It loses points on conciseness due to heavy restatement of the core invariant across sections, and on progressive disclosure because the referenced companion files are absent and their would-be content is inlined.

Suggestions

Deduplicate the snapshot-before/verify-after invariant and the ax-vs-px escalation ladder: state each once in the behavior-matrix/ladder section and reference it elsewhere instead of re-explaining.

Create the referenced companion files (MACOS.md, WINDOWS.md, LINUX.md, BROWSER.md, RECORDING.md) and move the inlined per-platform and feature-specific material into them so SKILL.md is a true overview.

Tighten or relocate the repeated action-fact and route-vocabulary explanations; they reappear under 'Read action facts' and again in the behavior matrix.

Dimension	Reasoning	Score
Conciseness	Mostly efficient operational prose, but the snapshot-before/verify-after invariant and the ax-vs-px escalation ladder are restated across four-plus sections (behavior matrix, ladder algorithm, 'Verify after every action', canonical loop, dispatch table), so it could be tightened substantially by deduplicating.	3 / 5
Actionability	Provides fully executable, copy-paste-ready guidance: concrete bash/JSON invocations, a complete tool dispatch table mapping intent to tool calls, and a working PIL crosshair script covering the common cases.	5 / 5
Workflow Clarity	The canonical loop and the verify-then-escalate ladder give a clearly sequenced algorithm with explicit validation checkpoints (verify_state satisfied/unknown) and validate→fix→retry feedback loops, including explicit cautions for destructive and batch operations.	5 / 5
Progressive Disclosure	References to one-level-deep companions (MACOS/WINDOWS/LINUX/BROWSER/RECORDING.md) are clearly signaled, but those files do not exist in the bundle and ~1050 lines of per-platform/feature content that belongs in them is inlined into SKILL.md, leaving the structure partly aspirational.	3 / 5
	Total	16 / 20 Passed

Description

92%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A strong, third-person description that concretely names the capability (snapshot/act/verify across multiple addressing modes) and gives an explicit Use-when trigger clause with natural keywords. Minor synonym coverage gaps keep trigger term quality just short of perfect.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions — 'snapshot its accessibility tree, act through snapshot-bound element tokens, native menu paths, exact window geometry, or pixel coordinates, and verify from fresh state' — giving comprehensive, specific coverage rather than abstract language.	5 / 5
Completeness	Explicitly answers both what (drive a GUI app via CLI/MCP, snapshot, act, verify from fresh state) and when ('Use when the user asks you to operate, drive, automate, or perform a GUI task in a real application on the host') with concrete trigger phrases.	5 / 5
Trigger Term Quality	Includes natural phrases a user would say ('operate, drive, automate, or perform a GUI task in a real application') plus 'native GUI app', but a few common synonyms (e.g. 'control my app', 'click the UI') are not covered, so it sits just below the comprehensive anchor.	4 / 5
Distinctiveness Conflict Risk	Occupies a clear niche — native GUI automation via the Qwen Cua Driver CLI/MCP across macOS/Windows/Linux — with distinct triggers and minimal overlap with other skills.	5 / 5
	Total	19 / 20 Passed

Validation

75%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 12 / 16 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (1082 lines); consider splitting into references/ and linking	Warning
metadata_version	'metadata.version' is missing	Warning
metadata_field	'metadata' should map string keys to string values	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	12 / 16 Passed

Repository: QwenLM/qwen-code
Path: packages/cua-driver/rust/Skills/cua-driver/SKILL.md
Commit: a858dea

Reviewed: 4 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.