I’m terrible at spotting flaws in my own work. As I iterate through multiple drafts and careful editing, I still tend to lose myself in the details and focus more on the trees than the forest. It’s not uncommon for me to miss weak arguments, unclear explanations, and assumptions that need defending—until after I publish, of course.
While preparing my previous post about AI writing workflows, I decided to test something: asking Claude to actively look for problems with my draft from an antagonistic perspective. Not just gentle feedback, but the kind of opposition you’d face from readers with strong opposing viewpoints.
The results were more brutal than I’ve been used to from Claude, yet they were more useful, than I expected. Here’s what I learned by using adversarial review to stress-test arguments, and why this technique shows promise despite some significant limitations.
The experiment setup
Close to publishing, I gave Claude this prompt:
Review the most recent draft as an antagonistic reader who would like to find any and all problems with the article (whether constructive or just to be antagonistic). This reader might have a pro-AI agenda to push or an anti-AI (as in no AI can be good) agenda.I was looking for the kind of opposition that finds every possible angle of attack—the readers who come to your work already convinced you’re wrong and looking for ammunition to prove it.
Claude responded by creating and applying three distinct antagonistic personas:
- The ideological opponent (challenges core assumptions and claims bias)
- The pedantic critic (nitpicks definitions and demands citations)
- The bad faith reader (misrepresents positions for rhetorical advantage)
Then it proceeded to tear into my article from both pro-AI and anti-AI perspectives. I had asked for this, and the responses were deliciously brutal. They were antagonistic in exactly the way I’d requested, and to be honest, genuinely useful.
What different types of opposition reveal
Claude organized the feedback by its perspective on AI. What follows is a summary of the feedback. You can see all the feedback in the transcript of the chat.
Pro-AI critique: Spotting unclear boundaries
The imaginary AI enthusiast hit me with accusations of “gatekeeping” and “elitism”:
This whole hand-wringing about ‘integrity’ is just fear-mongering that holds back progress. You’re basically saying AI can’t be trusted to help with learning, but then you admit you use it extensively in your own writing. Classic gatekeeping.
My initial reaction was to note how this missed my actual argument—this felt like a shallow reading that cherry-picked inflammatory words. But the critique exposed something important: apparently, I hadn’t clearly explained what I meant by AI tools needing supervision for learning new topics. The boundary between “AI can help” and “AI shouldn’t be left alone” was muddier than I realized.
The “guardrails” criticism was similar:
Your ‘guardrails’ argument is elitist – you’re essentially saying only experts should use AI tools. That’s like saying only mechanics should drive cars because they can spot problems.
Again, my first instinct was to point out how this missed my actual argument. But it revealed that I’d failed to explain why AI writing assistance differs from other tools that democratize complex tasks. I believe that GPT-tool users require sufficient subject-matter knowledge to evaluate responses because the GPT-tool responses can be misleading.
I think the mechanic-car example is quite good because early cars really needed to be operated by people with considerable mechanical knowledge, Eventually, cars became more reliable and automobile use became more standardized, reducing that requirement, while adding requirements for training and licensing. There are a lot of parallels to the current state of AI tools that can be drawn here.
Anti-AI critique: Exposing unaddressed assumptions
The anti-AI persona proved more effective at finding real weak spots. Instead of misreading my arguments, it attacked assumptions that I hadn’t defended:
You’re part of the problem. This entire article is an elaborate justification for using plagiarism machines while pretending to maintain ‘integrity.’ You admit you can’t tell if AI content is original or stolen, yet you use it anyway.
This comment identified something I’d completely avoided: the question of where AI training data comes from. I’d written about workflow integrity while sidestepping the broader ethical questions about the content these tools were trained on. Something that shouldn’t be forgotten in a conversation about the integrity and ethics of using AI tools.
The “automated cheating” critique was equally pointed:
Your workflow is just automated cheating with extra steps. Step 1 claims you preserve ‘thinking work,’ but Steps 2-4 show AI fundamentally reshaping your ideas and expression. How is that still ‘your’ writing?
Unlike the pro-AI critiques, this one engaged with the substance of my argument. It forced me to confront a question I’d been avoiding: at what point does AI assistance cross the line from editing tool to ghost writer?
(Full disclosure: Claude says this article is still at least 85% my content. But I’m not sure where the line between editor and ghost writer is.)
Methodology insights from this experiment
The anti-AI feedback proved more valuable than the pro-AI responses, and I think I understand why. The pro-AI critiques attacked positions I hadn’t taken, while the anti-AI critiques attacked positions I had taken but hadn’t adequately defended.
This suggests something important about constructing effective antagonistic prompts: opposition works best when it assumes the worst possible interpretation of your actual arguments, not when it misrepresents those arguments entirely.
I’ll have to try having the persona steelman the paper’s arguments by presenting the points in their strongest form before refuting them. This could help clarify and then further strengthen the articles points before they are published.
The goal of refining the personas and prompts it to inform the AI what I’m looking for as specifically as possible while giving it the freedom to find what I haven’t thought of.
Current limitations and open questions
Even in its currently crude form, this technique revealed gaps in my reasoning. The antagonistic review identified weak spots without providing the expertise needed to strengthen them. Even if the AI can’t fix weaknesses it finds, finding them is still helpful. By identifying gaps in the article, it identifies gaps in my understanding. In the spirit of our collaboration, this shows me where I still need to do some more homework before I can call that the article is ready.
The process also depends heavily on the AI’s ability to generate credible opposition. Claude’s responses felt plausible, but I’m not sure they represent the strongest possible critiques a real expert might offer. This suggests that how well it works might depend on the subject matter. For example, if the GPT doesn’t know much about the topic, such as might be the case in technical topics that the GPT hasn’t been exposed to, it might not be able to present an informed opposition. Even worse, it might not know be able to tell you that unless it’s been prompted.
There’s also the question of when to apply this technique. Too early in the writing process, and it’s stress-testing arguments that aren’t fully formed yet. Too late, and you might discover fundamental problems when there’s no time to address them, although, I’d argue that it’s better late than never, even in this case. But finding the sweet spot is important.
What I’m testing next
The three-persona approach (ideological opponent, pedantic critic, bad faith reader) provides a useful framework, but I want to refine the prompts and develop the personas. Instead of asking for “antagonistic” readers, I’m experimenting with requesting “the strongest possible objections” to specific claims.
I’m also curious whether this works better for certain types of content. Opinion pieces and argumentative writing seem like natural fits, but what about explanatory articles or technical documentation?
Practical takeaways
Antagonistic review won’t replace subject matter expertise or catch structural problems, but it’s useful for stress-testing arguments before publication. The technique seems to work best when you:
- Apply it to nearly final drafts, not early exploratory writing
- Ask for opposition to positions actually taken in the article, not strawman versions
- Focus on one or two personas rather than trying to cover every possible angle
- Use the feedback to identify weak spots, then bring in the appropriate expertise to address them
The goal isn’t to elicit every possible criticism. Instead, it’s to find the stinging critiques that point directly to where the reasoning needs work. For the cost of a single prompt, it can save you from publishing arguments that aren’t as strong as you thought they were.
That seems like a reasonable tradeoff.