Getting started¶
This article explains how to install the Qt skills into your AI tool of choice, invoke the skills, and verify that the skills are actually used.
1. Install¶
To install agent skills to your AI tool:
Claude Code automatically discovers the skills, so no restart needed. The Claude Code plugin install also wires up the Qt Documentation MCP server in the same step.
Restart Codex after installation. MCP setup is separate — see Manual setup.
See the project README for manual install, VSCode agents setup, and per-skill symlinking options.
2. Invoke your first skill¶
Open a project containing Qt code, then ask the agent to do
something a skill is designed for. Trigger phrases are
defined in each skill's description — the agent matches your
request against them.
Examples to try (substitute a real file path):
"Review my QML changes in
Main.qmlbefore I commit.""Document this class —
src/network/RequestQueue.h.""Profile this app and find what's making the UI feel laggy:
build/myapp.exe"
The agent will load the matching skill, follow its workflow, and report back.
3. Verify it actually fired¶
Skills load on demand. A request that sounds relevant doesn't guarantee activation. Look for these signals:
- Claude Code — a
Skilltool invocation appears in the transcript with the skill name, for exampleqt-development-skills:qt-qml-review. Without it, the model answered from general knowledge. - Codex CLI — the skill appears in the loaded-skills list at session start, and the agent quotes its workflow steps.
- Gemini CLI — the extension is listed under
gemini extensions listand the agent references@SKILL.mdcontext.
If the right skill didn't fire, see Why a skill didn't activate.
4. Verify the MCP server¶
Ask the agent something only fresh Qt 6.11 documentation can answer:
"Using the Qt 6.11 docs, what's the default value of
QQuickWindow::persistentGraphics?"
A working MCP server returns a qt_documentation_search tool
call and a precise answer with a doc link. A missing server
results in either a hedged guess or a refusal to consult docs.
Per-client troubleshooting lives in MCP → Verifying.