Generate Arm Performix AI insights for a Code Hotspots run

Select a specific run

You’re now ready to generate AI insights using Performix and Claude Code.

Use a run ID whenever possible. A run ID prevents Claude Code from selecting a similarly named run or a more recent run from a different workload.

If you don’t have the ID, ask Claude Code to list supported runs:

    

        
        
Use Arm Performix to list supported Code Hotspots runs. Include the run ID, target, workload, and creation time for each run.

    

Generate AI insights for a specific run

Request an insight for the selected run, replacing <run-id> with the ID returned by Arm Performix:

    

        
        
Use Arm Performix to generate an AI insight for run ID "<run-id>". Identify the highest-impact finding, cite the profile evidence that supports it, and suggest the first investigation or optimization step. State any missing evidence or uncertainty.

    

If you’re investigating a specific function, make the request more focused:

    

        
        
Use Arm Performix to explain why function "<function-name>" is hot in run ID "<run-id>".

    

Focused prompts usually produce more useful results than broad application-wide prompts.

Evaluate and refine AI insights

A useful AI insight identifies the run that it analyzed and connects its conclusion to measured evidence. The evidence can include sample percentages, call paths, source attribution, or disassembly. An insight should also state when the available data can’t establish a root cause.

If a recommendation isn’t traceable, ask for one finding and its evidence:

    

        
        
Focus on the highest-impact finding in run ID "<run-id>". Cite the measured evidence, distinguish hypotheses from observations, and identify the next Performix view or recipe to inspect.

    

You can ask a similar question for a known hotspot:

    

        
        
For function "<function-name>" in run ID "<run-id>", summarize what Code Hotspots proves, what it doesn't prove, and which additional evidence is needed to identify the root cause.

    

The following are some other questions that you can ask:

    

        
        
Which measured evidence in this Performix run supports that recommendation?

What alternative explanations fit the same evidence?

Which additional Performix recipe or view would distinguish between those explanations?

What is the first optimization I should try, and why?

    

Code Hotspots locates sampled CPU time. If you need evidence about microarchitecture, instruction mix, or memory access, use a more specific Performix recipe.

Iterate and measure

After you review the evidence, use the following controlled optimization loop:

1. Record the baseline workload, target, run configuration, and elapsed time.
2. Change one performance-relevant variable.
3. Rebuild the workload.
4. Rerun the same workload on the same target.
5. Profile it again with the same Performix settings.
6. Compare elapsed time and profile evidence with the baseline.

Claude Code can help with more of this loop if your environment exposes build, deployment, remote-execution, or source-control tools. Review proposed code changes and command approvals before Claude Code applies or runs them.

What you’ve accomplished and what’s next

You’ve now generated an AI insight and reviewed its supporting evidence.

Next, you’ll open the same run in Arm Performix to inspect the underlying profile views.