Ai Tricks

How I Actually Use AI Without Getting Burned

Most people overcomplicate this. They run a prompt, get a generic wall of text, and assume it's useless. That's because they're treating AI like a word factory instead of a thinking surface. Here's what actually works if you're willing to do some real work instead of hitting "generate" and calling it done. The biggest mistake I see is people expecting raw AI output to be publication-ready. It isn't. Nobody is. The trick isn't finding a magic prompt—it's building a workflow where the AI does the tedious stuff and you handle the judgment calls.

Ai Tricks That Actually Save Time

Start with structure, not prose. Feed the AI an outline, bullet points, or even just a few key facts you need covered, then ask it to expand. This gives it guardrails. When you let it free-associate from a blank page, you get exactly what you'd expect: plausible-sounding but hollow content that reads like everything else online. I keep a personal style reference. Not a document I revisit every time—just a few past pieces I wrote that I'm happy with. When the AI starts sounding too much like a textbook, I paste one of those in and say something like "rewrite this to match this voice." Takes about ten seconds and immediately grounds the output. The part nobody talks about is iterative refinement. You don't get it right on the first pass. I usually run three passes minimum: one for raw content, one for rewriting with better flow, and one specifically for removing the things that sound robotic. Repetitive transitions, excessive hedging, that overuse of "furthermore" and "additionally"—those are dead giveaways.

Specific Techniques I Use Regularly

Context stacking is the closest thing to a genuine shortcut. Instead of asking the AI to guess your audience, tone, and purpose, you give it all of it upfront. "You're writing for small business owners who know nothing about taxes. Tone should be conversational but not casual. Avoid jargon or explain it when you use it. Length: roughly 800 words." You get dramatically better output the first time around. Wasted less time adjusting. Reverse outlining comes in handy when you have a decent draft but it feels scattered. Paste your AI-generated draft back into the model and ask it to extract the outline. What you usually find is that your piece has five main points when it should have three, or the logic jumps around in ways you didn't catch while writing. Fix the structure at the outline stage. Fixing it paragraph by paragraph is way more painful. I also use AI for thing it's unexpectedly good at: finding gaps. Paste your finished draft and ask "what would a skeptical reader object to here?" or "where is the weakest evidence?" The model will point out holes you actually didn't notice. This isn't reliable for deep expertise—don't ask it to fact-check technical claims—but for catching obvious weak spots in reasoning, it's useful.

Where It Falls Apart

Here's the honest part. AI struggles with genuinely novel thinking. If you're trying to develop an original argument, a new framework, or anything that requires connecting dots that don't already have a clear path between them, the model will give you the most probable answer, not the most interesting one. Probability is not the same as insight. It also hallucinates citations with enough confidence to make them look real. I learned this the hard way. I was working on an article about conversion rate optimization and cited a "2023 Baymard Institute study on checkout form optimization" that the AI generated. It sounded legitimate. The study didn't exist. I had to scrap the whole piece and rewrite it from scratch. Now I verify every single source, even the ones that look perfectly reasonable. Took me two hours I could have saved if I'd just been skeptical from the start. Another failure mode: recency. Models trained on older data don't know about developments that happened recently. If your topic involves current events, new regulations, or recent research, the output will include outdated information as if it's still accurate. I always cross-reference dates against the current year before using anything.

My Actual Workflow

Here's what a typical session looks like for me: I spend 15-20 minutes gathering my own notes, sources, and structure before I touch the AI. This anchors the output in real knowledge rather than training data probability distributions. I feed the AI my structured notes and ask for a rough draft. This usually takes 3-5 minutes of generation time and gives me something to work with.

I rewrite the draft myself, keeping maybe 40% of the AI's words and replacing the rest with my own voice and judgment. This is where the actual work happens. I run the result through the gap-finding prompt I described earlier, fix the issues, then read it aloud one more time. Reading aloud catches awkward phrasing that your eyes will skip over when reading silently. The whole process takes about 45 minutes for something that would have taken me 2-3 hours from scratch. The AI isn't doing the thinking. It's doing the heavy lifting on the drafting part so I can focus on the parts that actually matter: accuracy, voice, and whether the thing is worth reading.

Quick Reference

If you want tools, I use ChatGPT for the drafting and brainstorming, Claude for longer-form analysis where context window depth matters, and Perplexity when I need to verify facts against current sources. No single model does everything well. That's normal. The trick is matching the tool to the task instead of forcing one into every situation. Bookmark this if you find it useful. I'll update it as the tools change, which they will—probably sooner than you expect.