Operators Using AI
Practical Stuff

Pitfalls & Things to Set Up Now

I'm not an AI safety expert. These are just the things I've learned the hard way, and the settings and habits that save me time every day.

Always Get Citations

AI makes things up sometimes. The fix is simple: tell it to always cite its sources. Put this in your user profile (Settings → Profile) so it applies to every conversation.

Always provide citations and sources for factual claims. If you're not sure about something, say so. Never make up statistics, studies, or quotes.

And when you get output with facts in it, follow up:

Which of these claims are you most and least confident about? Flag anything you're assuming or that I should verify independently.

AI Slop Is Real (and Your Bar Should Be Higher)

I want everyone using AI. If you're not, you're leaving tools on the table. But here's the thing: once I know you have these tools, my expectations go up. If you have access to Claude and Perplexity and your memo still has no data, no research, and filler language everywhere, that's a problem.

AI slop is when someone clearly ran something through AI and hit send without thinking about it. Generic phrasing, no specifics, reads like a template. You can spot it in two seconds. Don't be that person.

The standard now: proposals should have more research behind them. Memos should have data. Emails should be tighter, not longer. AI gives you capacity to do better work, not just faster work. Use it that way.

Review this [memo/proposal/email] I drafted. Is any of this generic filler that doesn't say anything specific? Flag sentences that a reader would skim past. Make every sentence earn its place.

Ask Claude What You Can Share

Not sure if something is okay to paste into Claude? Just ask it. Seriously — Claude knows what's sensitive and can tell you what to redact before you share it.

I want to share [type of document/data] with you to work on it. Before I do, what should I redact or anonymize? What's safe to include and what's not? Assume I'm on a paid plan and my company doesn't have a formal AI policy yet.

Quick baseline: Claude Pro doesn't train on your conversations by default. Don't upload customer PII (emails, phone numbers, financial data) unless your company has approved it. And if your company doesn't have an AI usage policy yet, flag it — it shows you're thinking about this responsibly.

Use Voice Dictation

On Mac: turn on dictation (System Settings → Keyboard → Dictation). Press the microphone key or double-tap the fn key, and start talking. It transcribes in real time. This is a game changer for getting thoughts into Claude quickly.

Instead of typing a long prompt, just talk. "Hey Claude, I need to put together a project plan for the Q2 vendor review, here's what I'm thinking..." Your spoken thoughts are often better context than what you'd bother typing out. Then let Claude clean it up.

Kill the AI Voice in Your Writing

If everything you send starts sounding the same (slightly formal, weirdly enthusiastic, full of "leverage" and "streamline"), you've outsourced your voice. Put this in your profile:

Never use these words or phrases: leverage, streamline, delve, dive in, game-changer, I hope this email finds you well, please don't hesitate to reach out. Write like a real person. Short sentences. No filler.

Better yet, paste in 2-3 examples of emails you've written that you liked. Tell Claude "this is my voice, match it."