BD screen-vision
AI screen vision and desktop computer control skill for OpenClaw. Let your AI agent see the screen, understand UI elements, and autonomously perform mouse and keyboard operations (click, type, scroll, drag) via a screenshot-analyze-action loop. Cross-platform: Linux (headless server with XFCE4+noVNC, or desktop), macOS (cliclick), Windows (pyautogui). Supports any OpenAI-compatible vision API (SiliconFlow, OpenAI, DashScope, Zhipu, Ollama, etc.). Smart diff detection saves tokens. Safety mechanisms block dangerous operations. Trigger: when user asks to operate/control computer, view/interact with screen, open applications, browse websites, fill forms, perform desktop GUI tasks, take screenshots, or any task requiring visual screen understanding and desktop automation. Keywords: computer use, screen control, desktop automation, GUI agent, visual agent, 屏幕操控, 桌面控制, 视觉代理, 截屏, 自动化操作, 远程桌面, AI操控电脑, screen vision, desktop agent, mouse keyboard automation, UI automation, computer use agent, CUA, computer control, screen interaction. Examples: "打开Chrome搜索天气", "看看屏幕上有什么", "帮我操作电脑", "截个屏", "open browser and search weather", "click the submit button", "take a screenshot", "fill this form", "帮我打开微信", "操作电脑下载文件", "远程操作桌面".
As a process D 44/100 · Unfinished process — weak spots: result and completion, when it triggers, inputs and preconditions
What is at stake
Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The skill contains commands that delete files, rewrite disks or execute code fetched from the network. The agent may run them without asking if it believes the instructions require it.
Replace destructive commands with safe equivalents that ask for confirmation, scope them to one folder, and stop piping curl into a shell: pin a version and a checksum.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".
If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.
How to improve
- Shorten the description to 1024 characters.
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 3
✓ No critical or high findings
Medium and low: 3
-
medium Dangerous commands
cmd-execpolicy-bypassREADME.md:150Runs PowerShell with execution policy bypassedpowershell -ExecutionPolicy Bypass -File scripts/setup/setu…ps1
-
medium Dangerous commands
cmd-eval-dynamicscripts/core/run_task.py:220Dynamic code execution from decoded/untrusted inputos.system(f"cp {ss_path} {log_ss} 2>/dev/null") -
medium Exfiltration
net-redirectable-api-keyscripts/vision/analyze.py:182Helper sends the API key to a host configured by an environment variable — the key can be redirected to another serverAPI key + configurable base URL from environment
Files scanned: 19. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- error
description-longdescription is 1230 chars, limit 1024
Process rating: all ten parameters 44/100
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 0Progress reporting. Says nothing while it works
- 20When it triggers. No condition that starts the skill
- 30Running it twice. 2 mutating operations with no state check
- 40Consistency. Frontmatter name (screen-vision) differs from the folder (ai-screen-vision)
- 50Failures and branches. 0 branches, has a failure section
- 60Tools and files. Uses tools (bash, python) that frontmatter does not declare
- 100Steps. 17 steps
- 100Execution cost. Instruction body is 1003 tokens
- medium Safety rules and hard prohibitions inside a skill: they belong in the system prompt, here they protect nothing
- low The response is described with custom markup (3 tags): a typed call is more reliable
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +4Description does not say when NOT to use the skill (false activations)
- +3Description length 1229: 120–800 characters recommended
- +3Output format is not stated: the model decides each time
- -220 emoji in the instructions: noise for the model
- +1No license
- +2Single-language instructions
- +5Description quotes 10 example trigger phrases
- +4Structure: 13 headings
- +3Step-by-step instructions: 17 items
- +4Has examples (3 code blocks)
- +4Reference files are cited in the instructions (3 of 3)
Quality base 70; lint remarks subtract, signals add up to 100. Result: 63.