Bundle archaeology beats "make it faster": 2.1 MB to 890 KB, one change per measurement

Before it had read a byte of build output, the agent proposed swapping a library that accounted for 4 KB ^[8]. Set that against the 1,258 KB that eventually came off ^[1] and the suggestion was worth about 0.3 percent of the job ^[3]. The rest of the cold advice was worse: lazy-loading three components that were already lazy-loaded, and considering tree-shaking that was already on ^[7]. That is what an agent returns when the only input is source code, and it is the strongest argument in the post for building a reporting step first.

The more expensive lesson was the batch. Five optimisations applied together took 300 KB off and silently stopped two charts rendering, with no way to identify the culprit short of unwinding all five ^[12]. Roughly a quarter of the eventual win ^[4] arrived in a shape that had to be discarded. After that the author wrote the loop into CLAUDE.md as a hard rule: record the size, change exactly one thing, re-run size and tests, commit with before and after in the message or revert on the spot ^[14][13].

This is one practitioner's self-report on dev.to ^[18], and it is worth being precise about what the numbers say. The starting figure is given twice, 2,148 KB raw and 612 KB gzipped; the finishing figure is given once, as 890 KB, with no compression state attached ^[19]. Raw against raw, that is a 58.6 percent cut ^[2]. If the original 28.5 percent ratio ^[6] still held at the end, 890 KB is roughly 254 KB over the wire ^[7], but the post does not supply that number and no reader of a bundle-size result should have to infer it.

The transferable artifact is not the percentage. It is the report script: a walk over the dist output and the generated source maps that emits a flat map of package to bytes contributed ^[9]. That is what turned a vague brief into a checkable one. The follow-up prompt shows the delta, since asking for every import site of the top ten packages returned file paths and line numbers, plus three packages imported at app root but reachable only from /reports ^[10][11]. The precondition is a production build that emits source maps for something to read.

Four sessions for 1,258 KB is about 315 KB a session ^[5], on a dashboard that had accumulated for three years and carried 214 direct dependencies ^[17][4]. The reading was the cheap part. What got the work funded was 11 percent of mobile sessions bouncing before first interaction ^[5], not the Lighthouse score of 41 ^[2] and not the 8.4 second time to interactive on a throttled Moto G4 profile ^[3]. Lint rules and a CI budget are credited with keeping the result from rotting ^[15], though the protocol text available here breaks off before either is shown.