By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres walks through her experience vibe coding production software over six months, from early failures to a reliable workflow. She explains what vibe coding is, compares popular apps and agents, and identifies the 'doom loop' where agents keep failing to fix bugs because requirements shifted and code layers fell out of sync. Her remedy is two review cycles: plan-review-fix to clarify intent in markdown before coding, and implement-review-fix to catch agent mistakes in error handling, tests, and security. The piece matters because it offers a practical discipline for product people building real software with AI agents. The full article is partly paywalled, with transcripts and skills reserved for subscribers.
01Key takeaways
- Plan in markdown with the agent before writing any code, so mistakes are cheap to fix.
- Have a separate reviewer agent evaluate the plan from a fresh perspective, and iterate until you, the planner, and the reviewer are all satisfied.
- Start new conversations frequently to avoid context rot and degraded agent output.
- Use a code reviewer to specifically check error handling, unit and integration test coverage, and security basics like secrets and input validation.
- When a bug resists fixing, diagnose with a fresh agent first and only allow code changes once the cause is confirmed.
02Key sections
- Vibe coding defined and its tool landscape
- Vibe coding means describing software in natural language and letting an AI write it, a term popularized in early 2025. The author separates no-code-style apps that handle hosting and data from agents that offer more control but require infrastructure management.
- The vibe coding doom loop
- Projects go wrong when requirements are unclear or constantly changing, leaving stale code and mismatched data, controller, and view layers. Agents then fail to fix bugs despite claiming success, and the author's dog health tracker shows how starting over can succeed.
- Plan-review-fix cycle
- Iterate on a plan in markdown rather than in code, then have a separate reviewer agent check it for gaps, over-engineering, and blind spots without fixing it. Start fresh conversations often to avoid context rot and question suggestions you don't understand.
- Implement-review-fix cycle
- After implementation in a fresh session, an AI code reviewer scrutinizes bugs, duplication, over-engineering, error handling, test coverage, and security. The human stays in the loop to judge the reviewer's findings rather than letting agents iterate endlessly.
- Debugging by separating diagnosis from fixing
- When an agent struggles, start a new conversation and ask for diagnosis only, sometimes with several agents to see if they converge. Only let code change once the root cause is confirmed.
03From the post
“I've written more software in the past six months than I've written in my entire lifetime. I started vibe coding in March 2025. My first foray was extraordinarily successful and then resulted in complete failure. I was blown away when Replit helped me build a custom chat interface”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.