Claude Code's whole-repo claim rests on eleven modules and 2,300 lines

Eleven modules and 2,300 lines average about 209 lines a module ^[7]. A project that size fits inside a single prompt, so the whole-repo claim here is a context-window fit. The hard case in a big codebase is deciding what not to read, and a 2,300-line project never poses it. The run does show call-graph tracing across all eleven modules plus one circular import the author had never noticed, in a single turn ^[5].

His comparison numbers are softer than that. "Cursor starts to struggle past six or seven files. Copilot gives up around two," he wrote ^[6]. Those are thresholds from six months of paid use, and the post does not describe running the same refactor through all three tools ^[22]. He pays for all three himself and keeps all three installed ^[21].

The scope statement is the part worth copying into your own evaluation. "If you build similar back-end, script-first work, this review will transfer cleanly. If you're writing React all day, your mileage will differ," he wrote ^[4]. The work in the sample is Python finance automation glued to broker APIs and Google Sheets, running on cron jobs or Railway containers ^[3].

Build time is 38 percent below his pre-Claude baseline ^[16]. Debugging time is up slightly, "because I now take on projects I wouldn't have before" ^[18]. Both lines come out of the same consulting log. If the project mix changed, the 38 percent covers a different population of work than the baseline it is measured against. Which way that cuts, he does not say.

Broker-specific knowledge is where the adoption cost sits. On the Zerodha Kite Connect SDK he puts it at 80 percent right, with F&O ticker formats, post-2024 margin changes and specific error codes in the missing fifth ^[11]. For one client it used a parameter name that had been deprecated nine months earlier, and the test suite caught it before production ^[12]. The wrong 20 percent is vendor-specific and plausible. Reading the diff does not catch that; a test that calls the API does.

Cycling has no automatic detector in this loop. After three failed fixes the agent sometimes tries approach A, moves to B, then returns to A wrapped in a helper, so he stops it by hand, pastes the stack trace into a fresh context and names the constraint, as in "do not change the signature of X" ^[13].

The post prices a big refactor as "the equivalent of a cheap dinner" ^[14], with no per-session token figure. A dinner is not a per-session cost. His controls are bounded budgets, smaller models for simple sub-tasks, and a ceiling on every agentic loop, the same techniques he used to cut a client's API bill from 85,000 rupees to 12,000 rupees a month ^[15].

The revenue line is capacity: 40,000 to 70,000 rupees a month of extra billable capacity ^[19]. Held every month across the six, that is 240,000 to 420,000 rupees ^[20]. Forty automations in that window is close to seven a month ^[23]. For the 38 percent to transfer you would need Python scripts of a few thousand lines, one reviewer who knows the domain, and tests that already exercise the vendor SDK.