Skills for profiling and documentation

[richardkmichael]

@pauldruce referred me to the toolkit from a comment I made on the MATLAB MCP Server repo.

For several months, our team has been using two MATLAB skills I wrote: one for profiling, another for docs. I thought I'd share them.

Our profile skill predates matlab-optimize-memory and matlab-optimize-performance in the toolkit. I designed it to handle more general user queries, without a priori awareness of the workload, e.g., "Why is it slow?", or "Why is ABC taking so long?". The agent can figure out which factor is relevant: CPU, memory, I/O, call frequency, etc.

I know the toolkit doesn't accept pull requests, but I used a fork of the toolkit to add evals and structure the skills similar to those in the toolkit. I thought this would make it easier to try them if anyone is interested. The front-matter is Claude-focused, but should be easily generalized.

https://github.com/richardkmichael/matlab-agentic-toolkit/tree/profile-and-docs-skills

I'd be happy for feedback and to possibly upstream them into the toolkit.

Details:

The matlab-profile skill teaches the agent to:

• use the MATLAB profiler to capture function calls, time and memory usage, etc.
• iterate on its own, rerunning the profiler as the findings demand
• explore alternatives by writing and benchmarking experimental scripts

User code is not directly modified; the agent makes recommendations based on findings. The natural next step is to direct the agent to apply the recommendations.

Profile info is serialized as JSON for querying by the agent directly or via bundled scripts, and also saved as HTML for human review.

Profile runs are stored, for comparison with past or future runs.

The bookkeeping of the profile runs is partly delegated to the agent via skill instructions. I didn't want to write and include a full-blown "catalog management" tool. This could be improved, especially if the agent behaves poorly at maintenance.

The matlab-docs skill teaches the agent to:

• read local MATLAB docs
• use / search mathworks.com docs, if local docs are not installed