A free worksheet for checking a slow browser task before choosing a fix

[liubrain39]

I maintain BrowseSprint, an independent research project, and made a free worksheet for checking one slow browser task before deciding what to change.

The available resources are a public diagnostic guide, a blank six-run CSV, and an editable comparison protocol. They are usable without signup. This is a manual learning resource, not an extension or automatic timing service.

For the exercise, choose a repeatable, read-only task on a public page with a result you can check:

1. Write the expected result and source URL before starting. Decide what evidence would establish completion.
2. Record three baseline attempts. Keep failures, timeouts, retries and the actual session conditions.
3. Change one factor that addresses the observed problem, then record three changed attempts under comparable conditions.
4. Compare correctness and elapsed time together. Leave measurements you could not observe blank; an incorrect answer that arrives sooner is not an improvement.

I also published result-verification field notes, with a standard-library Python check of public HTML. The examples distinguish a form reset from a receipt, multiple anchors on one page from multiple outcomes, and a submitted proposal from accepted inclusion. They are author-run observations, not a controlled benchmark.

Three runs per condition are a small diagnostic exercise and cannot establish a general speed claim. Browser state, network variation and changes to the page may still affect the comparison. Total elapsed time alone does not identify an internal model bottleneck.

I'd value feedback on whether the success rule and session-condition fields are clear enough for a repeatable public-page task.

Disclosure: these are my project's resources; this post was prepared with AI assistance. No measured speed gain is claimed, and BrowseSprint is not affiliated with OpenAI.

[TheodorNEngoy]

I’d distinguish a fresh Codex conversation from a fresh browser tab/session in the session fields. Those can vary independently, and a reused conversation may already contain the answer to a repeated lookup. Recording the model and reasoning setting, when visible, would help too.

The success rule is clear. An optional observed_result or evidence field would make success=true easier to check alongside the expected outcome. I ran into this distinction while checking Turnfeed in Codex: response time and a verified result were separate questions.

I’ve read the protocol, but haven’t run the six trials.