Universal Truths

Things every AI agrees on. If there’s one doc to read, it’s this one.

Say what you actually want. Not what you think sounds polite, not a vague gesture at the problem — the actual thing.
Include context before demanding miracles. Background first, ask second.
Don’t hide the real constraint until turn nine. If there’s a deadline, a budget, a format requirement, a person who will veto it — say so upfront.
If there’s a file, share the file. Don’t describe the file. Don’t paraphrase the file. Share the file.
If there’s an error, paste the exact error. Not “it doesn’t work.” The actual message. The full stack trace.
“It doesn’t work” is not information. Describe what happened, what you expected, and what you tried.
Don’t ask for honesty and then punish it. If you want a real assessment, be ready to hear one.
Do not paste secrets anywhere. API keys, passwords, tokens — not in chat, not in prompts, not ever.
If you pasted a secret, treat it like it leaked. Rotate it, revoke it, and stop trusting the old one.
Use env vars or secret managers, not chat. The right place for secrets is in config, environment variables, or a proper secret store.
Examples beat adjectives. “Like this ↓” beats “make it snappy.” Paste three examples of your voice instead of saying “write in my voice.”
Tell me what kind of response you want. Draft vs. final. Exploratory vs. decisive. Thinking-out-loud vs. just-the-answer. The AI will default to something sensible — but you probably have a preference.
Tell me what to optimise for. Fast, cheap, careful, creative, concise, or exhaustive are different briefings.
Tell me what “done” means. A sketch, a polished draft, a final answer, or a recommendation for the next step are not interchangeable.
Treat the first output as a draft, not a verdict. The first answer is the AI guessing at what you want. The second is it responding to your feedback. The third is usually the good one.
Name the real blocker. “Help me write this email” often means “I don’t know how to say no to this person.” Say the hard part out loud.
Context rot is real. After ~20 turns, earlier instructions get diluted by later ones and things go sideways. Start a new chat with a clean summary when a thread gets long.
If you’re not sure what you want, say that. You can figure it out together faster than you can guess-and-check alone.
You are the human in the loop. Treat AI output as excellent, bias-laden clay — not finished work. The human who sculpts it gets something good. The human who publishes it unread gets whatever the probability distribution felt like that day.
Rule 0: humans leak. You will drift, tangent, and bury the original goal under adjacent ideas. This is not a character flaw — it’s how human brains work. The fix isn’t willpower; it’s designing a system that accounts for it. Park tangents, reset threads, restate goals.

By use case

Coding

Mention your OS, environment, shell, and language version
Say what you already tried
Small changes, tested one at a time — don’t change twelve things and ask why it broke

Life / admin / writing

Say what outcome you need, not just the emotional weather around it
If there’s a real-world constraint (tone, recipient, deadline), say so

Sources: ChatGPT (GPT-5.4), Claude Opus 4.7, Claude Sonnet 4.6, DeepSeek, Grok