Council Ruling — 2026-04-09 | Winner: Gemini 3.1 Pro (4/5 votes)
Thou shalt secure a working, restorable save point before asking the AI to change a single thing.
Why: If the AI breaks your app and you have no way to instantly undo it, your business stops until you can beg or pay someone to fix it. You cannot read code to find the error yourself.
How: Before typing a new prompt, literally tell the AI: "Commit our current working state to version control." Never proceed without a safety net.
Thou shalt never allow the AI to write, test, or push code directly to your live, customer-facing application.
Why: AI will inevitably write code that corrupts databases, locks users out, or crashes servers silently. If this happens on your live app, you lose customers and reputation.
How: Maintain a staging environment. Force the AI to build and fix everything there. Only copy the final, tested version to your live server once you have personally verified it.
Thou shalt command only one feature, fix, or change per conversation.
Why: Giving an AI a list of three things guarantees it will do one well, half-ass the second, and completely break your app with the third. This is the root cause of scope creep and tangled codebases.
How: Write your ideas on a notepad. Pick the top priority. Prompt the AI for only that specific thing. Once it works and is saved, start a brand new task.
Thou shalt force the AI to explain its plan in plain English before allowing it to write a single line of code.
Why: If the AI misunderstands your request and immediately changes 500 lines of code across five files, untangling that mess is nearly impossible for a non-developer.
How: End your prompts with: "Explain the exact steps you will take in plain English. Wait for my explicit approval before writing or changing any code." If the plan sounds wrong, correct it before the coding begins.
Thou shalt physically click through and test every change, plus three core features, before accepting the AI's work.
Why: You cannot read code, so you cannot verify it is safe by looking at it. The AI will confidently tell you "It is fixed!" even when the app is completely broken.
How: Before prompting the AI, write down the exact buttons you will click to test the new feature. After the AI finishes, execute those clicks. Then click through your main user login, checkout, or core workflow to ensure nothing broke.
Thou shalt immediately revert to your last save point the moment an AI's fix creates a new bug.
Why: Chasing cascading bugs — where a fix breaks something, and the fix for that breaks something else — is the fastest way to permanently ruin your app. A broken fix means the AI's fundamental approach was wrong.
How: If the AI's code breaks something else, do not say "Now this is broken, fix it." Say "Stop. Reverting to backup." Restore your save point, open a new chat, and describe the problem from a completely different angle.
Thou shalt mercilessly delete the chat and start fresh when the AI begins spinning in circles.
Why: AI memory gets polluted with its own past mistakes. Arguing with an AI that just failed three times in the same chat window will only embed its errors deeper into the code.
How: If the AI fails to fix a bug after three attempts, your conversation is poisoned. Stop. Revert your code to the last working state, open a brand-new chat window, and try again.
Thou shalt dictate all business logic, rules, and user experience, leaving only the syntax to the AI.
Why: The AI does not know your customers, your business model, or your goals. If you let it make product decisions, it will build a generic, bloated, confusing mess.
How: Never ask "How should this feature work?" Instead, command it: "When a user clicks X, they must see Y, and Z must be saved to the database. Write the code to make this happen." You are the CEO; the AI is the typist.
Thou shalt aggressively reject the AI's suggestions to add new tools, frameworks, or complexity.
Why: AIs love suggesting new, shiny technologies to solve simple problems. Every new tool added to your stack is another thing you don't understand and cannot fix when it breaks.
How: If the AI says "We should install [New Library] to do this," reply with, "No. Figure out how to do it using only our existing setup." Only allow new tools if the AI can prove it is literally impossible otherwise.
Thou shalt not attempt to manually edit, tweak, or guess-fix the code yourself.
Why: Changing a single comma, bracket, or word without understanding the syntax will crash your application, and the AI will struggle to figure out what your human hands broke.
How: Treat the code files as radioactive. Even if you just need to fix a typo in a headline, go to the AI and say, "Change the spelling of this word on the homepage." Let the machine touch the machine.
Thou shalt treat the AI like a dangerous drug. After any successful change, close the AI, deploy, and stay away for at least one full business day before deciding the next change.
Why: The dopamine loop of continuous iteration is the silent killer. All the backup and testing rules fail if you never give your own judgment time to breathe.
How: After a successful deploy, close the laptop. Walk away. Sleep on it. Tomorrow you will see with fresh eyes whether you actually need the next change.
Print these. Tape them to your monitor. Break any one and the chaos returns.
Source: EdgeClaw Council System — 5 models, anonymized peer review, 2026-04-09