Every developer using AI tools has hit this at least once.
You ask Cursor or Copilot to fix one bug. It touches 8 files. Three of them weren't related to the bug. One of them broke something else. You spend the next hour undoing what the AI "fixed."
This isn't an AI problem. It's a workflow problem.
The AI had no idea what you wanted from it. Fix surgically? Refactor freely? Plan first? It didn't know. So it guessed. And it guessed wrong.
The root cause
Most developers use the same prompt workflow for every task.
Bug fix. New feature. System design. Quick patch. Same structure. Same expectations. Same results.
The AI treats all of these identically because you treated them identically. You gave it no operating context.
The fix: Operating Modes
I built Vibe Coding Essentials to solve this for myself. The core idea is simple.
Before prompting, set a mode. The mode tells the AI exactly how to behave for this specific task.
4 modes:
[BUILDER] — New features. Move fast. Acceptable tradeoffs. Don't over-engineer.
[MAINTAINER] — Surgical fixes only. Touch nothing outside the scope. No refactors. No new dependencies.
[ARCHITECT] — Plan and structure before a single line of code is written.
[ECONOMY] — Minimal output. Low token use. No padding. No unnecessary files.
Usage is simple:
[MAINTAINER] We have a regression in the billing calculator. Fix only the calculateTotal function. Write a failing test first.
[BUILDER] Build a user authentication flow with email and password. JWT tokens. Don't add OAuth yet.
[ARCHITECT] Design the database schema for a multi-tenant SaaS. No code yet. Just decisions and tradeoffs.
How to get started
60 seconds. No install. No dependencies.
Copy AGENTS.md to your project root
Paste the anti-hallucination rules for your framework into AGENTS.md
Add a mode tag before your next prompt
See if the output changes
That's it.
Why I built this
I got burned. Multiple times.
Context collapse mid-feature. AI adding 4 packages to format a date. A session that cost real tokens and shipped nothing. Code that worked but nobody on the team could understand.
I started documenting every failure. Patterns emerged. The modes came from that.
This is the framework I wish existed when I started using AI tools seriously.
The philosophy
The model generates. The developer decides. You own the workflow.
Contributions welcome — especially if you've been burned by an AI workflow failure that isn't documented yet.
