Detailed Analysis
A Reddit post detailing the 120-day build of Ticketmappr—a live-events aggregation app spanning web and Android—offers a granular, data-backed case study of what solo software development looks like when Claude Code (specifically Sonnet 4.6) is embedded as a primary collaborator. The developer, an 18-year industry veteran who had never previously shipped a side project to production, tracked every metric from git and session logs: 36,159 lines of code across Kotlin, TypeScript, CSS, and SQL, 286 files, 19 database migrations, and 86 commits over 49 active days. The transparency of the reporting—down to exact code-growth checkpoints and language breakdowns—lends unusual credibility to a genre of AI-productivity claims that often rely on vague anecdote rather than verifiable numbers.
The most substantive insight isn't raw output speed but a shift in the economics of software iteration. The developer notes that 28% of all commits were corrections, removals, or full rebuilds of previously shipped work—including five separate rebuilds of the onboarding flow and at least three ground-up architecture rewrites. Historically, sunk-cost bias keeps developers wedded to suboptimal early decisions because throwing away working code feels expensive in time and ego. With an AI coding agent absorbing much of the implementation burden, that psychological tax drops sharply, making deletion and rearchitecture a routine part of the workflow rather than a rare, painful event. The tradeoff, as the author candidly notes, is a new failure mode: frictionless addition means codebases can accumulate cruft unless developers deliberately prune, since the agent will happily keep building on top of anything.
Equally notable is where the developer's own expertise remained indispensable. Problems like budgeting infrastructure capacity between live query serving and background import jobs, or rebuilding a batch write path that broke under real data volume, required domain judgment that prompting alone couldn't produce. This matches a broader pattern emerging across reports of AI-assisted engineering: large language models excel at generating and iterating on code quickly, but architectural tradeoffs, capacity planning, and knowing which technical decisions are load-bearing versus provisional still depend on human systems-thinking. The eight CLAUDE.md context files (1,735 lines total) the developer maintained per codebase area also point to an increasingly common practice—curating persistent, structured context to keep an AI agent aligned with project-specific constraints across long-running, multi-session work.
The broader significance lies in what this represents for the "unshipped side project" phenomenon common among experienced developers—the graveyard of prototypes abandoned before reaching production because the last mile of polish, deployment, and platform-specific work (in this case, a six-week sprint from first Android commit to Play Store release) demands disproportionate effort relative to the initial prototype. By compressing that effort, agentic coding tools appear to be lowering the activation energy required to finish, not just start, software projects. As Anthropic continues to position Claude Code as a tool for sustained, multi-session, multi-file engineering work rather than isolated snippet generation, case studies like this—with their session logs, transcript sizes (up to 1,000+ turns in a single 16MB session), and commit-level granularity—serve as real-world evidence for how such tools are reshaping the economics of solo and small-team software development.
Read original article →