中文指南 | Deutscher Leitfaden
Software Coach is a floating desktop assistant for learning and operating complex software step by step. It does not click buttons for you. It looks at your screenshot, combines that with your question, and returns structured guidance.
- explains what screen you are probably looking at
- suggests the next step to take
- explains why that step matters
- shows confidence, assumptions, and unknowns
- supports progressive coaching with
I completed this stepandStill stuck
The first built-in software pack is PerfView.
- Launch
Software Coach. - Press
Ctrl+Shift+Spaceto show or hide the floating window on Windows. - Choose a screenshot source:
Capture Screento hide the app and drag a region on the primary displaySelect Local Imageto use an existing image file
- Enter your goal or question.
- Click
Ask Coach.
Capture Screen hides Software Coach before capture. Then:
- drag a box over the part of the screen you want
- release the mouse to confirm
- press
Escto cancel
Current limitation:
- region capture works on the primary display only
Good examples:
CPU is high. What view am I in and what should I inspect next?Is this PerfView CPU Stacks or CallTree?I think this is a GC Heap issue. What should I do next?Does this screenshot match the expected Thread Time state?
If possible, tell the app:
- software name
- software version
- your goal
That reduces wrong guidance later.
Software Coach prefers to confirm software name and version before detailed coaching.
Behavior:
- if you provide software and version, the app uses them
- if you do not provide them, the app may ask you to confirm first
- if you still do not provide them, the app may guess from the first screenshot and ask you to confirm
This is intentional. It is safer to clarify context early than to coach the wrong software.
explain: explain the current UI or viewguide: give step-by-step operational guidancediagnose: reason about a possible problemverify: check whether the current screenshot matches an expected state
After a response, you may see:
I completed this stepStill stuckShow more detailWhy?Back
Use them to continue the same coaching session instead of starting over.
The app supports three provider types:
mockopenai-compatibleollama
Use mock when you want the app to run without any API key or model server.
Use this for cloud or local endpoints that expose an OpenAI-compatible API.
Typical fields:
Base URLModel Name
If you use a cloud provider, required API keys must come from environment variables such as OPENAI_API_KEY.
Use this when you run a local Ollama server.
Typical fields:
Ollama Base URL- installed model list
- custom model name
Current default local URL:
http://localhost:11434/v1
The app can:
- list installed models from
ollama list - let you choose one from the list
- let you type a custom model name
- if the model is missing, ask for confirmation before running
ollama pull <model> - show install progress in Settings
For software screenshot coaching, the current recommendation is:
qwen2.5vl:7b
Lower-resource fallback:
gemma3:4b
Larger alternative:
gemma3:12b
By default:
- screenshots stay local
- screenshots are not stored
- telemetry is off
- remote docs are off
- cloud models are off
What leaves your machine:
- nothing in
mockmode - nothing in local Ollama mode if Ollama runs on your machine
- screenshots and text may leave your machine only if you explicitly enable and select a cloud-capable provider
You can rate answers as helpful or not helpful. This feedback is stored locally to help future tuning.
- the first real software pack is PerfView
- screenshot region capture supports the primary display only
- no automatic clicking or keyboard control
- no overlay highlighting yet
- best results currently depend on a vision-capable model
- if confidence is low, the app will ask for more context instead of giving overconfident instructions
Current Windows shortcut:
Ctrl+Shift+Space
- make sure the main window is visible
- click
Capture Screen - if nothing appears, bring the app to the foreground and try again
Check:
ollama serveis running- the base URL is
http://localhost:11434/v1 - the selected model is installed
Try:
- giving software name and version explicitly
- taking a tighter screenshot of the relevant area
- asking a narrower question
For project architecture, development setup, testing, packaging, and extension details, see the repository developer docs:
- README.md
- README.zh-CN.md
- README.de.md