Whether you are reviewing Qt Quick or Qt Widget code with Qt’s agentic skills, you can reduce your token consumption significantly by upgrading to the latest versions while getting the same review quality.
We have been restructuring the QML and Qt C++ code review skills. We achieved major token savings by:
-
Running the AI deep analysis with a single agent instead of launching 6 separate subagents analysing the same code from different perspectives
-
Handling the AI deep analysis without any subagent based on the defined rules if the code-under-review is less than 2500 lines
-
Enabling a focussed code review only on code that has been modified during the AI session
When we created the qt-qml-review and qt-cpp-review skills, Claude Sonnet 4.6 and GPT 5.4 were the frontier models we worked with. At that time, enterprises enjoyed flat-fee subscription packages from GitHub and Anthropic. Hence, we built the review skills with speed and review depth in mind with 6 sub-agents. But times have changed and the change of billing methods towards token-based pricing, made it appealing to restructure the skills to be more token-efficient.
We tested the refactored skills and found that they have lost nothing in terms of review accuracy with the mid-tier latest models, such as Claude Sonnet 5.5, Gemini 3.8 Flash, and GPT 6-Luna compared to the results of the old and the updated skills. This is partially because deterministic linters play an important role in the overall results, but also that newer model architectures do need less “focus on a single review topic” to be successful.
When benchmarking the old vs the new skill on a small Qt Widget project (for example the Qt’s Diagramscene project) with Claude Code CLI and Sonnet 5.5, we noticed an average cost (over 5 test runs) dropping from a $1.25 to $0.31. Cost savings for diff-only reviews can be ever higher depending on the file size and the number of code changes.
Video: Code Review with Agentic Skill in Claude Code (Video sequences have been edited for clarity)
Meanwhile, we also implemented some other user feedback from the GitHub repo in these skills. So, keep on using them and let us know where we can do better. As with the previous skills, the deep analysis review quality depends on the agentic harness and the LLM in use.
You can find the updated skills from our GitHub repo or by installing the Anthropic qt-development plug-in. If you have installed the qt-development plug-in already in Claude Code CLI, then it will update automatically (unless you have stopped the plug-in synchronisation).
