A Month of Chat-Oriented Programming (checkeagle.com)

Nick Radcliffe spent a concentrated month “pair-programming” with Anthropic’s Claude Code (Sonnet 4.5) to revive an old Python App Engine project, producing roughly 20k new lines of code (the repo now totals ~23k Python lines) and 1,731 passing tests. In his setup Claude ran as a terminal app (npm-installed) in three modes—Default (request edits), Accept Edits, and Plan Mode (produce a plan with three approval options)—plus a permissive “--yolo” flag he avoided. Radcliffe reports he personally wrote under a hundred lines of the new code: Claude authored the vast majority. The workflow was highly formalized; he watched and intervened constantly, using git resets and immediate stops whenever Claude made questionable changes. The experiment shows chat-oriented programming (CHOP) can be effective today but only with heavy human oversight. Claude’s outputs are impressively broad and deep—drawn from massive public-code corpora—yet remain “library-like” knowledge rather than true understanding, producing both high-quality code and brittle, sometimes destructive hallucinations. Radcliffe likens the experience to SAE Level‑3 autonomy: productive but stressful because humans must stay alert to take over. Practical implications for the AI/ML community: LLMs are now viable, high-leverage coding partners for real projects if teams enforce strict SOPs, review diffs, maintain extensive tests, and accept operational costs (stress, energy, and governance).