So, here’s how I avoid that while still using AI to be a lot more efficient. It’s taken practice, but I have a process that takes full advantage of the tools without outsourcing my brain or my voice.

Step one: get it all out.

When my kid was in third grade, I noticed they were learning to write without stopping: get the whole idea out, without regard to spelling, grammar, or structure, then go back. I thought that was smart and it’s stuck with me since. The blank page has always tripped me up less than the reflex to fix every sentence the moment I write it.

I can think way faster than I can type, and I’m guessing that’s true for most of us. So, next time you start a writing project, try using the transcription feature. Open a new chat, click on the little microphone icon (or use a tool like Wispr Flow), and just start talking. Strange at first, but once you get used to talking it all out, ums and backtracks included, you’ll never go back. For me, it’s like brainstorming at a whiteboard with a pile of pens and sticky notes:

“I’m working on a paper for [whatever]. I’m going to ramble my thoughts on every angle of it, then I want you to help me organize them.”

Step two: define the story.

Now it’s time to refine what I’m trying to say. What point am I making? Why should the reader care, and what do I want them to come away with? Nothing new about this step, but I get here hours faster than I used to because I already have all my thoughts out in front of me. This is the most important part, though: if you’re letting AI decide what your point is, ask yourself how necessary the task really is.

What’s new and incredibly handy, though, is bouncing the story I want to tell off what I’ve provided to jumpstart my writing:

“The point I’m trying to make is X. My audience for this is Y. What I want them to come away with at the end is Z. Help me structure an outline with what I’ve already provided in support of that thesis.”

I’m asking AI for help again, but the help stays locked to my story. I’m not outsourcing my point to AI; I’m using it to assemble my whiteboard of sticky notes in support of my point, which I can then edit and adjust to get right.

Step three: pressure test.

This is the peer review before the peer review, and it’s low-key my favorite part because it’s something I couldn’t do before AI.

“I want to pressure test this thesis and outline against a group of critical personas. Up to five interested parties (my manager, a tech leader, a sales leader, an ops stakeholder, and one more I haven’t thought of) should review and tell me where they see holes, open questions, contradictions, missing data, etc. Write me a robust prompt to generate their review process and tell me how each should be executed.”

This prompt is doing a lot; if you haven’t done meta-prompting, it may take a beat. I’m prompting the tool to generate up to five prompts I’ll run against my thesis and outline from five different perspectives. I then run these prompts against my thesis and outline in separate chats (sometimes in different AI tools) to get a diverse set of feedback. It doesn’t replace my critical eye; it adds information I probably wouldn’t have thought of in a fraction of the time. It shows me where to tune the story, suggests additional research, and sharpens my point before an actual peer review. It’s how I clean up the holes in my game before asking peers to spend their time on it.

Step four: draft, reclaim and refine.

Last step: pull my outline, the persona feedback I want to keep, and anything else I’ve added into a first draft. I ask AI to make this first cut, which sounds like outsourcing, and it is, but it’s outsourcing the mechanics, built from my story, my outline, and additional data that I’ve scrutinized. I know what I’m looking for. I’ve also built a writing style skill: explicit instructions (adjectives and adverbs sparingly; rely on data; never hyperbole, etc.), learned from ingesting 100% my own hand-written content, plus “AI-cleansing” filters and rules against inventing content (not foolproof, but it helps). Want to build one in about 30 seconds? Paste this into your favorite AI tool and attach some of your best writing as a sample set (Word, PDF, blog link, whatever):

“I want to build a skill that will explicitly apply my personal writing style to any content as well as remove typical AI patterns from written content, e.g. ‘The honest truth.’ Build me a robust prompt that reverse engineers what you know and can infer about my writing style from chat history and the attached sample documents. Include filters for typical AI bad habits including writing style and content influence.”

You’ll get a surprisingly good first draft of a reusable skill. Mine took months of tuning to get where it is, but this gets you 80% of the way in one paste, and you’ll refine it every time you use it.

After that, it’s all me. I drop the draft into Word and go through it with a fine-tooth comb. There’s not a single word, number, or semicolon going out the door that I haven’t inspected. Then I run the final version back through the tool: “Note the changes I made and remember those for future content generation.” I’m teaching the tool to work how I want it to work.

So, I’m not drafting anything important in minutes, but I’m definitely cutting days down to hours. It’s the difference between letting the tool do my thinking vs using it like a partner. I hand over the mechanics but keep the story, the judgment, and the final word on every sentence. That’s how I get more done, faster, without adding to the pile of AI slop, or publishing something I don’t want to stand behind.

Reflecting on the last 13 years of my career at Amazon, I pulled up my history of annual reviews. I have about 10 years of feedback available all in one place, including manager and peer comments, strengths, and opportunities for growth. I took the whole bundle and ran it through Quick looking for trends over the last decade. Did I have any opportunities for growth in 2013 that are now strengths? What growth opportunities persist today? How had my direct reports’ feedback on my leadership changed over time? It was validating to see some areas, like people management, where I took specific steps to improve and the feedback validated the change over time. Other areas, like holding just a beat too long to make some decisions (I occasionally marinate in data for longer than I should), remain frustratingly present.

I’m not sure I would have been able to pull meaningful results from roughly 260 paragraphs of narrative feedback across a decade without an AI assistant. Maybe if I spent an uninterrupted Saturday on it. Maybe. I’m 100% sure, though, that no tool or amount of time could have made sense of it unless the feedback was 100% human-generated. We weren’t using AI tools in 2013, obviously. Or 2014, or 2017, or even 2022, so nearly all the feedback was written by dozens and dozens of people who took the time to write down their thoughts for no other reason than to coach me. Looking at 2024 and 2025, though, as chat bots and agents began to creep their way into every facet of “knowledge work,” I could see some of the feedback starting to flatten. A few entries had that “AI sheen” to them. You know what I mean. Don’t get me wrong; I’m not judging. I experiment all the time with AI to do my job better, faster, all of it. When annual review time comes around and I have around 20 requests for feedback, it will be an obvious temptation to rip through the whole stack and be done with it in an hour instead of my usual “write three of these a day for a week” process. But this exercise was a clear reminder that even in a tech-centric place like Amazon, some things just shouldn’t be outsourced. That feedback is so valuable. The review process should be a human process, even if it takes longer.

So, while it’s hardly a new sentiment to remember to keep the humans in the loop these days, this was a strong personal reminder for me. Generating 80% of an MBR with a well-tuned AI process is brilliant, and I’ll never go back. I’ll still use AI to help me in the annual review process, too, probably by culling through stacks of email and Slack to pull out details I wouldn’t have remembered on my own. But I think I’m going to stick with writing three reviews a day by hand. Human to human.

Originally posted on LinkedIn.