Lessons from running 12 companies' software built with Claude Code: verification habits over agent quality

Short summary

A non-technical business owner describes running 12 companies' software stacks built almost entirely through Claude Code over six months. The core lesson: the AI agent writes good code, but the absence of a verification habit is the real bottleneck. Three production incidents taught hard lessons about staging parity drift, silent write failures, and access-control gaps in rarely-visited routes — each now prevented by mechanical, repeatable audit procedures.

• •Staging clones drift silently; feature flags, pre-flight migration checks, and rehearsed rollbacks beat maintaining a 'safe' environment
• •Every write path must surface success or failure to the user — no silent 200s on failed writes
• •Access-control bugs hide in rarely-visited routes; only a full route × role sweep catches them

Generated with AI, which can make mistakes.