The Engineering of Intent, Chapter 27: Anti-Patterns — What Not to Do

This is Part 27 of a series walking through my book The Engineering of Intent. In the previous chapter, we covered the checklists I actually use. Chapter 27 is the mirror: twenty-five named anti-patterns — the tempting shortcuts that look like speed and cost you the quarter.


Read It Once All the Way Through. Then Come Back When Something Feels Off.

Each entry in the chapter names the anti-pattern, says why it’s tempting, and prescribes the discipline that prevents it. Here are ten of the twenty-five. The book has the full catalog with remedies.


Ten of the Twenty-Five

  1. “Let the Agent Decide.” Tempting because it feels efficient. Fatal because decisions are where intent lives; an agent deciding without you expresses its priors, not yours. The agent proposes; you decide. Always.
  2. “It Compiled, Ship It.” Compilation is necessary, not sufficient. Add types, tests, review.
  3. “One-Prompt-Solves-Everything.” Agents produce worse output when asked for too much at once. Break into increments.
  4. “Agent as Writer, Human as Reader.” Good software requires writing judgment, not just reading judgment. Don’t reduce your own writing practice.
  5. “Late-Night Ship.” The last 5% is where subtle judgment lives, and your judgment at 11 PM is worse than at 9 AM.
  6. “Trust the Model, Not the Code.” The model said so is not a basis for merge. Always read and run.
  7. “Rewrite From Scratch.” Existing code encodes years of bug fixes you cannot see. Refactor incrementally.
  8. “Solo Genius.” Solo mode is where bad decisions go un-audited. Subject hard decisions to at least one other human’s review.
  9. “Too Clever to Fail.” Elegance that is not robust to agent-generated variation is not robust. Prefer simple.
  10. “Forever Pilot.” Endless pilots are how companies talk about transformation without doing it. Set a graduation deadline.
💡 Three you may not have seen named before: “AI-First to a Fault” — sometimes a well-tuned regex beats prompting for one. “Exclusive Model Loyalty” — no model is best at everything; use at least two. “Meta-Meta-Meta” — at some point you must ship a thing the user uses; set a tool-building budget.

The Most Dangerous One

“Theory without practice rots into ideology. Reading about Vibe Coding is easier than doing it. Books like this one are maps; the territory is your editor. Do the work. Ship things. Fail. Adjust.”

Every anti-pattern in the chapter is, in some sense, a local optimum — it feels like speed, in the moment. The discipline is to notice that the local optimum isn’t the global one. The people who make Vibe Coding work long-term are the people who have, every day, resisted five or six plausible shortcuts.

âš  The two that interact dangerously: “Plan-Free Zone” (planning feels slow, so skip it) and “One-Prompt-Solves-Everything” (one mega-prompt feels efficient). Teams that fall into both together ship quickly for three weeks and catastrophically in the fourth. The planning-free, mega-prompt culture is where most AI-native rollouts visibly fail. It’s also the culture that’s easiest to slide into, because both shortcuts feel like speed.

Next up — Chapter 28: The Tips Archive — 100 Notes from Daily Practice. If Chapters 23 – 27 gave the named patterns and anti-patterns, Chapter 28 is the long-tail archive — the hundred smaller notes, heuristics, and micro-lessons I’ve been keeping. Pick and choose.


📖 Want the full picture?

The chapter gives all twenty-five anti-patterns with the “why it’s tempting” explanation and the “remedy” discipline. Read once cover-to-cover; flip back whenever the process feels off.

Get The Engineering of Intent on Amazon →

2026-05-13

Sho Shimoda

I share and organize what I’ve learned and experienced.