Audit Yourself to Get More From GenAI

More than a year into using generative AI daily, I wondered whether I was getting the most out of my AI use. There was no benchmark or feedback loop, and no one was grading my sessions with ChatGPT and Claude — until I created a self-audit.

I did what I’ve always done when faced with a process that lacked measurement. I studied every method I could find — prompting guides, conversations with colleagues, my own session patterns. I used AI to help me use AI better. Over time, I built a single self-audit prompt — one that encapsulates more than 30 habits for getting the most from AI.

Each time I ran the self-audit prompt, the output got sharper. The discipline became reflexive for me. That’s the real value of the self-audit: It made me better at using AI, in every session.

Now, at the end of any significant AI session, I simply prompt: “Review this session and assess it against my AI habits guide. Score how I did, identify what I missed, and guide me to apply missed habits.” Within a few minutes, I get a diagnostic that is uncomfortably specific about what I missed. I now have an answer to a key question: whether my process was good, not just the GenAI output.

A recent field experiment confirmed what I found through my experience. A research team that included MIT Sloan professor Jackson Lu randomly assigned 250 employees at a technology consulting firm in China to either use ChatGPT to assist with their work or to work without it.1 The employees with ChatGPT access were judged as significantly more creative by both their supervisors and outside evaluators. But the gains showed up exclusively among employees with strong metacognitive strategies — those who reflected on their own thinking, recognized knowledge gaps, and refined their approach when results were weak. That finding underscores that metacognition — thinking about your thinking — is the missing link between simply using AI and using it well.

AI widens the gap between disciplined and undisciplined professionals. People who skip the discipline generate more volume without more insight — a pattern consistent with what researchers at the University of California, Berkeley’s Haas School of Business called “unsustainable intensity” in findings published in early 2026.2

Knowing how to use AI is good — but to get the most value from the tool, you need to know whether you’re using it well. The self-audit gives you that.

A Self-Audit That Measures Five Key Goals

My self-audit prompt is organized across five goals: set up, refine, verify, own, and systematize. These goals represent a practice that experienced professionals have instinctively followed for years, long before generative AI’s arrival. You don’t need technical training to score well on this audit. You need to replicate the thinking and brainstorming process that you are likely already good at when conducting competitive research, responding to requests for proposals (RFPs), engaging in acquisition analysis, and planning a sales presentation, for example. It is your skill in the application of AI, not the AI itself, that makes the difference.

The self-audit assesses each generative AI session with five questions linked to each of the goals:

• Set up: Did you prepare the AI before asking it to work?
• Refine: Did you iterate on your own thinking, or just reprompt?
• Verify: Did you verify before trusting?
• Own: Did you make the output yours, or accept the default?
• Systematize: Did you build something reusable, or close the chat and start over?

You won’t score well on all five goals in every session — nor should you. But knowing which ones you missed, and why, enables you to change your next session. Think of it as AI holding a mirror to your own ability. It gets sharper every time you make it your own.

To illustrate what strong performance looks like at each goal, and what the self-audit is measuring, I applied the audit to an actual competitive due diligence analysis on a $5 billion global services company. Details have been modified for confidentiality, but the habits, prompts, and results are drawn from actual chat sessions. I’ll focus on the impact one goal at a time.

1. Set Up: Pass the Intern Test

What the self-audit measures: Did you prepare the AI with sufficient role, context, constraints, and materials before asking it to work — or did you jump straight to a question?

The most consequential decision in any AI interaction happens before the first prompt. It’s the decision to prepare.

I tell the AI who it should be, what it has to work with, and what I need it to produce. “You are an elite research analyst specializing in competitive intelligence. Here are the target company’s last two annual reports and its most recent earnings-call transcript. Assess this company’s ability to disrupt our core business within 18 months and recommend our strategic response.” That prompt will produce far better output than “Tell me about this competitor.”

I call this the “intern test.” If you handed your prompt to a brand-new hire with no context about your company, your industry, or your priorities, would they know what to do? If not, why would you expect your AI to?

Most readers will likely pass this test. Any GenAI prompting guide or video covers the basics of setup.

What gets overlooked is making clear what setup should not do — the negative constraint. I specify what I do not want: “Do not give me a generic SWOT. Do not hedge every statement. Do not define terms I already know.” And upload your materials. The more context you provide, the more accurate the output. It’s like telling a new team member “Figure out our competitive position” versus handing them your last three strategy decks and customer feedback.

Two additional practices make setup more effective. Before a significant AI chat, I run a preflight check: “What does a great outcome look like? What are the three most important things to get right?” After the first good draft, I generate a bridge summary so context carries forward, especially when I’ll be taking a long break between prompts or need to transition to a new chat. You might not have considered using this tactic before. A bridge summary is especially valuable if you tend to have long, multipart exchanges over days or even weeks. (In one case, Claude suggested doing so at time intervals to avoid having the conversation get too complicated.)

In the due diligence scenario, the difference in outputs before and after the self-audit was stark. While my first prompt was solid, the negative constraints and a preflight check were missing. The variable was me. What made the biggest difference? The negative constraint. Once I told the AI what not to do — no generic SWOT, no hedging, no defining terms I already know — the output became richer in insight and started reading like a briefing, not a book report.

2. Refine: Pass the Rethink Test

What the self-audit measures: Did you truly iterate on your own instructions and thinking, or did you simply reprompt for a better answer?

The first output from any AI session is a draft, not a deliverable. The real value comes from iteration. But the most productive iteration improves your own instructions, not the AI’s answer.

That’s metacognition in action. The person who pauses to ask, “What did I fail to specify? What assumption did the AI make that I should have preempted?” is exercising exactly the reflective discipline that separates high performers from the rest. AI rewards those who rethink their own instructions — not those who rephrase the same request.

I started catching my own patterns. Sometimes the output sounded right, but I couldn’t explain why — so I’d ask the AI to walk me through its reasoning, and the gaps would surface. Other times, I’d catch myself reprompting the same request with slightly different words and realize that the real problem was that I hadn’t broken the task down. The hardest one to admit: When I still couldn’t get what I wanted, it was usually because I couldn’t describe the desired goal clearly enough. Pasting in an example of output that showed what I was after worked better than trying to describe it.

One of the most powerful refining habits is embarrassingly simple: Ask the AI what you should be asking. “What question should I be asking that I am not currently asking?” That one prompt has produced more valuable insights than any other, in my experience.

When I applied these habits to the due diligence, they surfaced a critical insight I’d overlooked: The competitor’s employee sentiment data contradicted its public narrative of a thriving digital transformation. That disconnect between external messaging and internal reality changed my entire threat assessment. I never would have discovered that if I hadn’t challenged my own assumptions.

3. Verify: Pass the Trust Test

What the self-audit measures: Did you independently verify the AI’s claims, check its sources, and stress-test its confidence — or did you trust fluent output at face value?

AI output typically reads well — which can be a problem. It’s linguistically fluent and structurally polished, even when the underlying claims are fabricated, outdated, or mathematically wrong. This is a new kind of quality risk, and it misleads experienced professionals more often than they’d like to admit.

I once asked AI to summarize the regulatory history of the credit card industry, which I know well. The response was beautifully written, logically structured, and completely wrong on two key regulatory revisions. It read like an A-minus term paper from a student who’d skipped the reading. I almost didn’t catch it — because it sounded right. That’s what worried me. I knew the domain well, and I still nearly walked into a committee meeting with hallucinated data.

Since then, I’ve built verification into my routine. I ask the AI to surface and rank every assumption behind its answer. I request verifiable sources and note when the model can’t provide them. For anything involving numbers, I ask for step-by-step calculations. I’ve found two habits particularly effective: the temporal awareness check (“What is the date of the most recent information you’re drawing on?”) and the confidence stress test (“Rate your confidence in each factual claim as high, medium, or low”).

It’s the same discipline we’ve always followed: Verify before you trust; trust before you share.

During the due diligence, the AI flagged that its revenue figures were nine months old and rated its confidence in the regulatory settlement details as medium. When I verified the output independently, I discovered a $42 million enforcement action that the AI had understated. That single verification changed the risk profile of the entire analysis.

4. Own: Pass the Signature Test

What the self-audit measures: Did you actively impose your voice, your position, and your audience on the output — or did you accept AI’s default?

The real work starts here. I used to stop too early. Most of us do.

AI models default to hedged, tonally generic output. Left unguided, they produce content that is competent but indistinct — written by a smart person who seems to have an opinion about everything yet commits to nothing. That’s fine for a rough research summary, but it doesn’t reflect your voice or your style, and it’s not something you’d want to put your name on.

The first complete draft was exactly that: well organized, factually grounded, and thoroughly researched. But it was hedged throughout and read like a report designed to avoid being wrong rather than to help someone make a decision. When I forced the AI to take a clear position on the competitive threat, pushed it for unconventional strategic responses, and asked it to apply champion-challenger lenses, the analysis became richer and something I would stake my reputation on.

One technique I use at this stage is running a draft by a virtual personal board of directors that I built. These distinct personas help push my thinking and the AI’s analysis away from the default path toward the edges. I built AI-powered personas modeled on real personalities: v_SunTzu for power dynamics, v_Indra (Nooyi) for the human dimension, v_Mark (Cuban) for commercial realism, and v_Meg (Whitman) for operational rigor. What survives that gauntlet of virtual advisers is sharper and more defensible.

The habit most people underuse is calibrating AI to their own personality: how they think, how they argue, and what they won’t tolerate in a deliverable. Take ownership of the thinking, not just the editing. That’s when the output starts sounding like you.

5. Systematize: Pass the Reuse Test

What the self-audit measures: Did you build systems that make your next session better — or did you close the chat and leave yourself having to start from scratch next time?

Nearly everyone treats each AI session as a stand-alone thread — which may be productive in isolation, but the value doesn’t compound. Here, the discipline shifts from improving sessions to building systems.

Building repeatable processes out of one-off successes is what I do. Yet, early on in my GenAI use, I spent two hours building a detailed competitive analysis that delivered exceptional output — and then I closed the chat. I’d produced a great deliverable but captured none of the thinking that made it great. I should have known better. When I needed to run a similar analysis a month later, I had to start from scratch — the same role definition, the same constraints, the same verification steps, all rebuilt from memory.

Three habits make the difference. These are not habits you apply at the end of the conversation but throughout — after every prompt, at every logical checkpoint, or after a break.

First, maintain continuity. During any significant working session, I ask the AI to maintain a running summary of what we’ve accomplished, what’s still open, and what I will need to copy and paste to resume the conversation in another chat. This produces a bridge summary that makes it easy for you to pick up the discussion in a new session without losing continuity, especially if you run out of tokens on one chat.

Second, be a coeditor. Review the AI’s output after every prompt, or at logical break points, and feed your own judgment back in. You read what the AI produced. Some of it is good; some of it is wrong. Some of it is vague in ways you didn’t notice until you tried to use it. You fix it, mark it up, and hand it back: “Here’s my revised version. Use this as our new baseline and continue from here.”

Third, “templatize” what works. Every time you craft a session that produces exceptional output — a due diligence workflow, an RFP evaluation, a customer analysis — convert it into a reusable template. Replace the specifics with [variable] placeholders and save the session as what I call a macro-prompt — a single structured prompt that combines the entire session’s workflow so anyone can run it without having to start from scratch. Individual expertise becomes organizational capability.

That single due diligence session became a reusable macro-prompt I’ve now used for partnership evaluations, board position assessments, and acquisition analyses — each time just pasting it in the chat to start the conversation. From there, AI guides me step-by-step — instead of me guiding the AI — with all of the thinking intensity captured from the original session. After every use, I run a prompt to improve this macro-prompt for the next session.

How to Start Auditing and Improving

Below, I’ve shared the self-audit macro-prompt that includes all 30 habits to audit oneself. Think of it as a companion resource. You can just copy and paste it into an existing conversation you’ve been having with AI on a significant, extended topic. See what it tells you about your use of AI across all five goals and 30 habits. The self-audit will show you exactly where to refocus.

Then, start practicing these habits in your GenAI conversations wherever you see the opportunity.

Generative AI technology has already proved its capabilities and will keep getting better. The discipline is what unlocks real value — and that discipline will always be needed, regardless of which AI tool you use.

There’s one last thing I didn’t expect when I started this journey: The better I got at working with AI, the better I got at thinking without it.

Run the self-audit. See what it tells you about your critical thinking.

Apply these tips to get the most from the self-audit:

• Run it at the end of every significant AI session, not just occasionally. The habit of measuring is itself the discipline.
• Don’t stop at the scorecard. When the AI asks, “Would you like me to go back and apply the missed habits?” say yes. Then run the self-audit again. Repeat until you’re satisfied you’ve extracted the most value from the session.
• Track your scores over time. You’ll notice patterns — goals you consistently score well on and goals you consistently skip. Those patterns are your development road map.
• Improve the prompt itself. When the AI suggests improvements to this macro-prompt based on your session, review them and update your saved copy. The self-audit gets sharper each time you use it.
• Make it yours. Add habits that matter to your work, remove ones that don’t, or build in your own techniques. The 30 habits here are a starting point, not a ceiling.
• Share it with your team. When everyone runs the same self-audit, you build a shared language for AI session quality across the organization.