Six Laws for Talking to AI
I analyzed 192 AI coding sessions. The data showed 8,471 messages and 89 million tokens. The total cost was $518.
But cost per token is the wrong metric. I wanted to know how much of my communication was wasted.
I found that 60% of my sessions were forked from older ones. I was restarting conversations more than finishing them. I realized I have an execution latency problem, not a knowledge problem.
Here are the six laws to stop wasting time with AI:
Write rules immediately. I averaged 36 messages between knowing a rule and writing it down. If a correction applies to future sessions, write it in your config file now. The cost of writing is ten seconds. The cost of not writing is the rest of the session.
Control the blast radius. AI can break a whole site by changing one file. Before any operation that touches multiple files, force the AI to list every file and every change. Wait for your confirmation before it proceeds.
Finish your thought before typing. I spent too many messages refining rough ideas. Write your full specification in a text editor first. Include fields, constraints, and edge cases. Send the full spec once instead of seven micro-adjustments.
Eliminate vague instructions. Short messages like "change" or "do it" cost multiple clarification rounds. If a person with zero context cannot execute your instruction, add more detail. "Change" becomes "Add input sanitization to the form handler."
Look for patterns, not single fixes. When you find a bug in one file, do not just fix that file. Tell the AI to check all files in that category for the same issue. This prevents the same error from appearing in different places.
Define the scope clearly. If a constraint applies to everything, use the word "all." If you say "this function," the AI only touches that one function. Explicit scope saves you from repeated corrections.
The goal is not better prompt engineering. The goal is a process to stop repeating yourself.
Source: https://dev.to/antonio_zhu_e726fd856cd86/six-laws-for-talking-to-ai-4dan
Optional learning community: https://t.me/GyaanSetuAi