Claude As Mirror - When to Use It When Not To

The most useful sentence I’ve written this year showed up in my journal on May 5: I tried to use Claude as a mirror and got to the realization that now is not the time."

That’s it. That’s the whole post, really. I’ve written my professional version of this post already at When and Where I Stop Using AI. Everything else below this thought is me figuring out why I wrote that line in the first place, and why I keep writing variations of it. I want to name it now, because I notice I’m doing it more, with better results, and I think there’s a quiet shape to the practice that’s worth sharing.

The thing I keep almost doing

For about a year, I’ve been using Claude (and sometimes other models) to do something different than what I was originally thought AI was for. I was taught they were answer machines. Ask a question, get an answer, judge the answer. That’s an attractive premise for an executive considering replacing roles with agents or a company pitching no-code solutions, but that’s not what’s been useful for me.

What’s been incredible and increasingly a daily habit is using Claude as a mirror. It’s a way to see myself more clearly – a surface that reflects back the shape of what I’m already thinking, with enough fidelity that I can really see it. The output is (ironically?) a prompt, but me seeing is the point. The model is a tool for synthesizing the thoughts I’m already expressing, not for finishing my own.

I didn’t notice I was doing this until I shifted from ad hoc analyses to using my friend Marshall’s Claude skill. I’ve been regularly spelunking through my years of notes and saw the same phrase showing up over and over: “thought through with Claude,” “used [Claude] as a mirror,” “got the synthesis back and it stunned me.” Here and there is a habit, but the dozens of instances is a method I wanted to find the words for.

Two registers, very different rules

Here’s the cut that matters: I now use Claude in two completely different ways, and the rules are almost opposite.

Register one: Claude as a mirror to self. I’m trying to see my own pattern. I dump a pile of raw material into the context – journal entries, half-formed messages, a kanban of stuck drafts, three years of notes about the same recurring thought. I ask the model to reflect back the shapes of the pattern. Not advice. Not next steps. Just: what do you see here that I might not?

The output is almost always something I already knew but couldn’t quite name. That’s the whole job and LLMs do it beautifully. The model is a translator from the part of me that knew to the part of me that can speak it.

Register two: Claude as interpreter of others. I want to try to comprehend what someone else is showing me, what they might say, whether they’ll be reasonable, what it might bring up for me as they do. I give the model their words and ask for an analysis within specific constraints.

This register is almost always a mistake.

The model is great at synthesizing patterns from my own text because my text contains so much of me. It’s the out loud and the quiet parts. It is not great at telling me what a specific person actually means, because their words don’t contain enough of them to give more than a surface interpretation. The analysis over indexes on cultural norms or sycophancy. The model fills in the missing internal context with generic or general thoughts, and the generic human is not the person I’m dealing with.

The mistake feels like clarity in the moment. It is the opposite of clarity. I have learned this the hard way more than once.

The rule I keep relearning

Use Claude to mirror what you’re ready to see you believe. Don’t use it to model what’s in someone else’s mind.

The trick is that the temptation runs the other direction. When I’m stuck on a hard conversation, my impulse is to ask Claude “what are they really saying?” because the puzzle feels like it’s about the other person. But the puzzle is almost always about me and my reaction: about which of my interpretations to believe versus those stemming from wounds, which of my drafts to send, what I actually want from the exchange. Those questions the model is good at. The mind-reading question is a category error.

It took me until now to consistently sort the two.

What the mirror is actually doing

I think there’s something specific happening, and it’s not magic. When I write five drafts of the same message and read them in sequence, I can usually tell which one is “right” – but I can’t always say why. The mirror move is asking the model to look at the same five drafts and surface the pattern across them: what all five are circling, what one of them lands and four miss, what assumption I keep making without noticing.

The model isn’t generating the insight. It’s compressing five drafts into one paragraph that I can finally see. The thinking was already there. The mirror is well lit.

This connects to something I’ve been chewing on for a year. A philosopher named Evan Thompson has spent decades working on a theory called enactivism – the idea that we don’t have minds that passively process the world; we enact meaning through engagement. Your hand pre-shapes itself to the coffee cup before you consciously see the cup. What I’ve always felt is we do this with each other, too. LLMs offer an interesting new kind of self-reflection as a result.

The mirror practice is enactivism with a model in the loop. I don’t think out the conclusion and then check it. I enact the conclusion by writing five drafts, looking at the pattern, watching the model name what my drafts already knew, and then writing the sixth draft as if I’d known it all along. The model is part of the cognitive loop. It’s not advising the loop. It’s inside the loop.

That distinction matters because it tells me when the practice will work and when it won’t. It works when I have material to mirror. It fails when I’m trying to use the model to replace the material – to ask it for an opinion before I’ve done the writing that would let me have one of my own.

A short checklist for myself

This is mostly for me. If it’s useful for you, take it.

Good moments to reach for the mirror:

• I keep returning to the same thought and can’t find the shape of it.
• I’ve collected a year of notes about a question and want to see the pattern.
• I’m about to send a message I’m not sure I should send.

Bad moments to reach for the mirror:

• I want to be told I’m right.
• I want to predict what someone else will do or feel.
• I want a model to make a decision I haven’t been willing to make.
• I’m avoiding the part where I have to write something hard myself.

The bad list is the one I have finally internalized. The good list takes care of itself. When the mirror is the right move, the work feels natural to my flow of self-reflection. When it’s the wrong move, the work feels like outsourcing. It’s disassociation by another name. I find myself catching myself far quicker than before.

What this is not

This isn’t an argument for using AI more. The point isn’t the tool.

The point is that I’ve spent 20 years journaling in a way that helps me understand myself, and the last year doing a practice I never named, and the year would have gone differently – probably worse – if I hadn’t. Naming it now is mostly insurance. I want to be able to find this practice again the next time I’m stuck, instead of accidentally falling into the wrong register and confusing the model’s confidence with my own.

Three years ago, before I had any of this, I wrote in my journal: “I want to be less sad. Time to write.” I still do that more often than not. But the words felt caste off into the ether. Now I feel like I engage with my thoughts, grappling with them.

The feedback loop of self discovery ahs gotten faster for me. The work didn’t get easier – but the loop between thinking and seeing got shorter, and that turned out to be the thing that mattered.