Sitemap
The Speedcraft Lab

The Speedcraft Lab

Senior engineer sharing FAANG interview post-mortems, big-tech system teardowns, backend deep-dives, and honest AI-agent experiments.

Member-only story

Spotify Cut Claude Code’s File-Reading Tokens by 90% by Not Letting It Read Big Files

One engineer’s test: a 350-line hook sends big reads to Gemini Flash, at 10 to 30 seconds each

7 min readOct 21, 2026

--

Press enter or click to view image in full size
Big files go in, short summaries come out but a subtle bug can get lost along the way

Gemini 2.5 Flash read the code and came back with tidy bullets, the kind that look like a job done. Then came the catch: it “found surface-level patterns but missed a subtle thread-safety bug.” Claude Code, once given the right context, “spotted it in seconds.” That test, in Dimitri Mazmanov’s 3 September 2026 post on the Spotify Engineering blog, is why his routing keeps the cheaper model away from debugging.

The plugin is called shunt, and it exists to keep Claude away from big files. Mazmanov, a principal product manager on Portal by Spotify, reports that across four scenarios on one Java monorepo it cut bulk-read tokens by “around a whopping 90%”. That is one engineer’s four-scenario test of Claude’s bulk-read tokens, not a Spotify-wide production figure and not a bill.

The post publishes no before-and-after token counts, no per-scenario figures and no method for the mean, and it does not net Gemini’s cost against Claude’s saving. Readers get the percentage and nothing to check it against.

--

--

The Speedcraft Lab
The Speedcraft Lab

Written by The Speedcraft Lab

Senior engineer sharing FAANG interview post-mortems, big-tech system teardowns, backend deep-dives, and honest AI-agent experiments.