Most people who try AI end up using it for two or three things and then stop. Usually it’s writing assistance, maybe some image generation, possibly asking it questions they’d otherwise Google. That’s fine — but it leaves a lot on the table.
These are things I’ve actually found worth keeping around, not just interesting for one afternoon.
1. Turning a messy wall of notes into something navigable
You know that document — or that note, or that voice memo transcript — where you brain-dumped everything and now it’s completely unusable? Paste it into ChatGPT or Claude and ask it to pull out the main threads, group related ideas, and flag anything that seems contradictory or half-formed. Not “summarize this” — that just compresses the mess. Ask it to reorganize and surface tensions in your thinking.
I’ve used this after long calls where I took scattered notes, after reading a dense report, and after finishing a book I wanted to actually remember. The output isn’t always right, but it gives you something concrete to push back against, which is much faster than staring at the raw mess yourself.
2. Role-playing a difficult conversation before it happens
This one feels slightly strange the first time and then becomes hard to stop using. If you have a conversation coming up that you’re dreading — a performance review, a negotiation, telling a client their project is late — you can ask Claude or ChatGPT to play the other person and push back on you.
Give it some context: “Play my manager. I’m going to tell you I need to push back on this deadline. Be skeptical but not hostile.” Then actually do it. The model won’t perfectly simulate your specific manager, but it’ll surface objections you hadn’t thought through and force you to articulate your position out loud before it matters.
I prefer Claude for this. It holds a character reasonably well and doesn’t immediately cave when you say something confident. ChatGPT sometimes agrees too fast.
3. Auditing your own writing for assumptions you forgot to state
Here’s a more specific version of “proofread this”: paste in something you’ve written — an email, a proposal, a report section — and ask the model to list every assumption the reader would have to already know or accept for your argument to land. Not grammar, not flow. Just: what does the reader have to already believe for this to make sense?
This is surprisingly good at catching the places where you skipped a step because it was obvious to you. Obvious to you, invisible to the person reading it for the first time. I’ve caught real problems in client-facing documents this way, the kind where someone would have replied with a confused question instead of a yes.
4. Building a rough learning plan for something you know nothing about
Not “explain X to me” — that gives you a Wikipedia article. Instead: “I know nothing about X. I have about four hours total to learn enough to have an informed conversation about it with someone who’s an expert. What’s the order of concepts I should learn, what should I skip, and what are the two or three things experts in this field actively disagree about?”
That last part is key. The disagreements are where a subject gets interesting and where you learn fastest, because they force you to understand the underlying mechanics rather than just the surface consensus. Ask it for those first and work backwards.
I’ve done this for topics ranging from pension fund accounting to mycology. It’s not deep expertise — but it’s enough to ask useful questions, which is often all you need.
5. Generating a list of things that could go wrong with your plan
Pre-mortems are a real technique used in project management and decision-making — you imagine the project has already failed and work backwards to figure out why. AI is quietly excellent at this because it has no emotional investment in your plan succeeding.
Describe what you’re planning to do, then ask: “Assume this fails six months from now. Give me ten specific, plausible reasons why, ranked from most to least likely.” Not vague risks. Specific ones: “your supplier goes out of business,” “your pricing is too high for the audience you’re actually reaching,” “a free alternative launches two weeks after you do.”
You don’t have to address every risk it surfaces. But reading through a list of specific failure modes changes how you think about a plan in a way that a general risk review doesn’t. Some of those items will be wrong or irrelevant. Some will be uncomfortably on-target.
What these have in common
None of these are “AI does the work for you” use cases. They’re all closer to AI as a thinking partner — something that responds to what you’re actually trying to figure out, not just executes a task. The model is doing real work, but you’re still the one deciding what to do with the output.
That’s roughly where AI is actually useful right now for most people. Not autonomous agents, not replacing jobs — just compressing the time between “I have a fuzzy problem” and “I have something concrete to react to.”
One caveat: all of these work better if you give the model real context instead of a vague prompt. The pre-mortem exercise is useless if you describe your plan in one sentence. The conversation role-play is useless if the model doesn’t know enough about the situation to push back meaningfully. Quality of output scales pretty directly with how much useful context you put in. Annoying, but that’s how it is.
