Specify the slow interaction
Choose one slow operation such as initial load, search, scrolling or saving. Record input, data volume, device, network and cache conditions. A broad complaint that everything feels heavy is not a measurement boundary. Confirm repeatability before requesting a fix so later observations can be compared meaningfully.
Separate loading and runtime work
Loading and post-load interaction need different observations. Record the target journey with an appropriate performance tool and inspect network, execution and rendering work. Instrumentation can affect results, so keep capture methods consistent across the comparison. An editor agent is a reasoning aid here, not the measurement instrument.
Separate observations and hypotheses
Ask for hypotheses such as repeated computation or duplicate requests, then check them against the recording. Preserve observed slow intervals separately from untested explanations. Change one factor where practical so a result can be attributed. A plausible optimization list is not sufficient evidence to accept every proposed modification.
Protect behavior while remeasuring
A faster response must not come from stale data or omitted validation. Preserve behavior and measure the same interaction several times, including unusually slow runs. Distinguish noise from a sustained improvement. Less code or an added cache is an implementation choice, not by itself proof of better user-visible performance.
Example: investigate a slow list search
For a hypothetical list that slows while typing, fix the sample data and input sequence. Inspect duplicated requests or rendering work, then test one justified change. Recheck correct results, interaction responsiveness and clearing the filter. This is a measurement procedure, not an empirical speed claim about a named coding tool.
Preserve evidence for the next change
Preserve revision, capture conditions, recordings and untested scenarios. Traces can contain source or sensitive information, so review them before external sharing. Separate agent entitlement from profiling tools and ensure another maintainer can repeat the procedure. This creates a baseline for recognizing a later regression rather than relying on memory of how fast the application felt.
Official sources
Pricing, availability and feature limits can change. Check the linked official pages before subscribing.
- Chrome DevTools performance panelInformation checked: 2026-10-28
- Chrome performance recording referenceInformation checked: 2026-10-28
- Save and share performance tracesInformation checked: 2026-10-28
- JetBrains AI AssistantInformation checked: 2026-10-28
- AI in ZedInformation checked: 2026-10-28
- JetBrains AI plans and usageInformation checked: 2026-10-28
- JetBrains third-party and local modelsInformation checked: 2026-10-28
- Install JetBrains AI AssistantInformation checked: 2026-10-28
- Zed pricingInformation checked: 2026-10-28
- Zed external agentsInformation checked: 2026-10-28
- Zed local modelsInformation checked: 2026-10-28
- Cursor pricingInformation checked: 2026-10-26
- Cursor models and pricingInformation checked: 2026-10-26
- Cursor Agent overviewInformation checked: 2026-10-26