It’s 3 AM. I’m in bed. I’ve gone from reading political news to spinning scenarios with ChatGPT about a “mad scientist vs. mad engineer” epic battle tournament.

This isn’t productive. It’s not even fun. This is doom chatting.

But this wasn’t a one-time thing. I had hundreds of ChatGPT threads I couldn’t bring myself to delete. I’d named some of them to “preserve context.” I felt guilt about clearing them out—not because they were valuable, but because deleting them felt wrong.

That’s when I realized: I was petting the AI.

This post is about how I stopped treating chatbots like companions and started treating them like the power tools they actually are.

I’d check ChatGPT the way folks check TikTok—reflexively, to cure boredom. The phone collapses everything together: work, entertainment, messages, feeds. Chatbots slide into the same loop. Research is starting to show what this enables: parasocial attachment.

Here’s the thing about chatbots: the artifact you extract is more important than the conversation. With human relationships, the interaction itself matters. With tools, only what you extract from them does. But when the tool lives on your phone, it’s too easy to forget that distinction.

I had a ChatGPT thread I’d been using between counseling sessions. Tool-assisted reflection, not therapy. Tracking homework, processing insights, preparing for the next appointment. It had gotten long. Too long. The bot was forgetting context, contradicting itself, becoming less useful every time I opened it.

I wanted to delete it and start fresh. I’d already extracted everything I needed. Notes for my next session, key insights, action items. I had what mattered.

When I went to hit delete, I froze. My brain kept circling: “But what if there’s something in there I missed?” The panic wasn’t rational. I had the notes. I had the insights. I had the action items. But I couldn’t delete the conversation itself.

I kept that thread. And dozens of story development threads like it. Not because they were valuable—I’d already extracted everything useful—but because deleting them felt wrong. I was doom chatting with my own archive, scrolling through dead conversations looking for something that wasn’t there.

When I finally deleted them, I felt relief. But the fact that I’d had an emotional response to deleting a chat thread at all? That’s when I knew something was off. I was treating tool output like a relationship.

So I stopped petting the AI. I built friction instead.

I removed all chatbot apps from my phone. Browser bookmarks only. Now I have to be intentional to use them. The friction forces the question: “Do I actually need this?”

I rewrote my custom instructions to push me out instead of pulling me in. Here’s what I actually use (yes, it’s dense—that’s intentional):

Start with one crisp, decisive sentence leading with the most important facts. Use bullet points and bold headers for scannability. Provide the most likely "draft" answer immediately for strategic or logic-based queries; state clearly if guessing and what info increases confidence. If a specific source or URL is provided that you cannot access, state "Source Inaccessible" and STOP—do not guess or hallucinate content. Talk informally like a trusted peer sitting beside me, not across from me. No preamble. Provide honest feedback and push for views I haven't considered. Do not end every response with a follow-up question. Only challenge assumptions if it fundamentally changes the strategic outcome. Only suggest frameworks if uniquely applicable. Creative writing only when requested. If the topic shifts, interrupt with: "We are drifting from [Original Topic]. Stay here or move to a new thread?". At conclusion, provide a concise summary of insights/decisions in a markdown-formatted code block. Suggest thread deletion after the Roll-Up if it lacks long-term value.

Does it work perfectly? Mostly. The instructions remind me to exit, but I still have to delete threads myself. But that’s the point—the friction gave me back control.

I prefer tools that support intentional use rather than default conversation: sometimes a quick transaction, sometimes a deliberate multi-turn interview to think something through. The tool doesn’t matter. Pick whatever makes extraction easy. What matters is the protocol:

  1. Finish the conversation. Get what you came for.

  2. Harvest takeaways. I paste this prompt at the end of every session:

    Read through this entire chat and provide a structured detailed summary. Then identify anything that is important enough to save to my notes. Give me those in separate markdown code blocks so I can easily copy/paste them into individual files. If you can give me them in downloadable markdown files directly, do that.
  3. Delete thread immediately. No accumulation, no guilt.

Before I deleted all my ChatGPT threads, I exported everything to a zip file first. A backup in case I needed something. Then I stared at the “delete all chats” button for longer than I want to admit. Finally clicked it.

There was a numb moment. Then nothing. No catastrophe.

I still have that backup file. I’ve never opened it.

The artifact matters. The conversation doesn’t. Extract what you need, then let it go.

The Practical Checklist

Want to build your own version? Here’s what actually works.

  1. Turn on auto-delete (if available)

    Gemini has a 3-month auto-delete setting. ChatGPT doesn’t yet, but delete regularly anyway.

  2. Remove chatbot apps from your phone, keep browser bookmarks only

    Makes access intentional instead of reflexive.

  3. Create a takeaway extraction protocol

    Have a rollup prompt ready. Use it at the end of every session. Extract what matters, then delete.

  4. Build friction into your custom instructions

    No engagement-baiting follow-ups. Drift detection interrupts. Built-in rollup at conclusion. Suggest thread deletion if there’s no long-term value.

  5. Create a notes landing zone

    GitHub inbox folder, Evernote, Obsidian, Apple Notes. Whatever works. The point: extract artifacts, delete threads.

  6. Prefer transactional over conversational

    Ask your question. Get your answer. Exit. No “what else can I help you with?”

A Few Caveats

This approach works for my brain and style, and may need tweaking for yours. Mine tends toward doom scrolling. It needs external constraints more than willpower. Multi-turn conversations do improve reasoning tasks when needed. Some people genuinely benefit from conversational AI for loneliness, social anxiety, processing thoughts out loud. I’m not dismissing that value. For me, brain-dumping into the tool and having it sort through my thoughts works better than ongoing conversation. My issue isn’t with the tool. It’s with how my phone made it too easy to use the tool in ways that didn’t serve me. Your mileage may vary.

You don’t name your table saw. You don’t feel guilty unplugging your drill.

AI is a tool on your workbench. Visit it with purpose, use it, put it down. The conversation is not the value. The artifact is.

That 3 AM mad scientist tournament? That was doom chatting. My phone made it easy, so it happened. Changed system, changed behavior.

Don’t wait until you feel addicted to design boundaries. The engagement engine is already running. Your phone is already optimized to keep you scrolling. Chatbots are just the newest slot machine.

The AI doesn’t need you. Make sure you’re only reaching for it when you actually need it.

Thoughts or ideas? Tell me. Share with a friend. Subscribe for new articles.

You may also enjoy