Coding with Claude Code 2: You're Optimizing the Wrong Context (coding-with-ai.dev)

AI coding tools aren’t just judged by accuracy or context-window size — they’re judged by how they fit human attention. The author contrasts our obsession with giving models huge context windows (200K–2M tokens) and perfect prompts with the neglected cost of interrupting a developer’s flow. After A/B testing in August — switching from Claude to Codex powered by GPT-5 — they found Codex produced more accurate outputs but was slower, prompting frequent task-switching. When Claude Code 2 arrived, its responsiveness kept the author "in the problem," preserving working-memory graphs of file relationships, recent errors, and hypotheses, which improved overall productivity despite lower per-query accuracy. For the AI/ML community this reframes evaluation and product design: latency and the interaction rhythm matter as much as raw performance. Metrics should go beyond per-task accuracy to measure end-to-end developer throughput, time-to-resolution, and cognitive load. Practical implications include prioritizing responsiveness, session and state preservation, and UI features that reduce reload costs (focused panes, lightweight incremental responses, persistent in-context traces). In short, optimize not only machine context but human context too — protecting developer flow may yield bigger real-world gains than marginal accuracy improvements.