Author
Dani Reyes writes DevMoment field notes on shipping real software with AI coding agents. Present tense, one repo at a time.
A field note on Claude Code commands: which custom slash commands actually earned their place, which I deleted, and the 2026 question the listicles skip, when a command should have been a skill.
A month of running Cursor and Claude Code daily on the same Next.js project. What each one is actually for, where they bit me, the cost difference, and why I kept both instead of picking a winner.
Running three to five AI coding agents at once with git worktrees for two weeks: the exact setup, honest merge and cost numbers, the failure mode nobody warns you about, and when parallel agents are the wrong call.
One weekend, one invoice tracker, three AI app builders (Lovable, Bolt.new, Totalum). Honest field log: where each one shone, where each one stalled, and which one survived to Monday morning.
A 30-day field log of running an MCP server in production: four failure modes the tutorials skip, honest latency and cost numbers, and when an MCP server is the wrong tool.
Seven days of running Cursor background agents on a real Next.js side project. 23 tasks, 14 PRs, 9 merged, $11.40 spent. What shipped, what failed, the task spec that doubled the merge rate, and when to reach for an AI app builder instead.
I take a two-year-old expense-splitter app with a rotting stack and rebuild it from scratch in a single weekend. I scaffold the plumbing fast with an AI app builder, then spend my real hours on the settle-up logic and UX I actually care about. The takeaway: scaffold the boring parts, hand-build the soul, model the data before the UI, and ship the unglamorous last 20% the same weekend.
I wire an MCP server to my IDE's agent in about thirty minutes, and suddenly it reads my real Postgres schema and project files instead of hallucinating. MCP is just a standard way for agents to call external tools. I pick one server that solves a real annoyance, drop a small JSON config with command, args, and env, restart, and let the agent fetch its own context. That's the whole win.
Hour three of vibe coding is where the magic wears off and the work begins. The agent starts drifting from its own code, re-implementing helpers and calling renamed props. I stay productive by committing at every green moment, re-grounding the context against real files every 30 minutes, narrowing the agent's scope when it wanders, and starting a fresh session when the context gets poisoned. Flow is a tool, not the goal.
