Over the past several months my relationship with AI tools has shifted, and I think it’s a good time for an update.

I’ve discovered three key approaches that have significantly changed how I work with LLMs:

Voice transcription has become my preferred mode for complex thinking.

Multi-phase prompt development has replaced my tendency to create over-complicated instructions.

Multi-lens analysis has given me a more effective thought partner.

Voice Transcription

The most noticeable change has been incorporating voice transcription. I now use it for about 40% of my AI conversations, with the rest being direct text entry. What surprised me most wasn’t that it worked, but how well LLMs can parse the natural messiness of spoken conversation.

This discovery happened through the very process I’m using to help write this post. I knew I wanted to do an update about how I use AI today, but I wasn’t sure what, exactly, that meant. I had to think it through. So I decided to just start talking. I rambled, went on tangents, circled back to earlier points — all the things that happen in natural speech, and sure enough, I got to it all just by thinking out loud. Then, the AI came back with an organized, coherent analysis of my voice transcript that could use to start drafting. It was truly a sight to behold.

Voice transcription captures ideas and personal style in a way that careful typing often doesn’t. When I’m typing, I’m much more intentional about sentence structure and sequence. I do heavy self-editing as I go. When I’m talking, I have to keep moving forward. I can talk much more rapidly than I can type, and if a tangent comes up or something else occurs to me, I just add it to the conversation.

This momentum aspect has been huge. Voice transcription works particularly well for brainstorming, problem-solving discussions, and when I want to think out loud through involved ideas. I still prefer typing when I want precision (when I’m very clear about what I want as a result and need that level of control). For getting my thinking down and organized, voice transcription has become invaluable.

It’s also made my interactions much more conversational. My colleagues probably think I’m talking to myself a lot more these days, though, since I do this at my desk at work. Thinking out loud has always been a useful tool for processing ideas, and now I have an intelligent conversation partner on the other end. If you’d like to try it for yourself, check out the easy, free, and cross-platform Handy.

Multi-Phase Prompt Development

For any task I find myself returning to repeatedly, I’ve developed a structured multi-phase approach to creating prompts rather than starting from scratch each time.

Phase 1: Maximalist Development

I start by asking the AI to help develop an exhaustive prompt. I explicitly encourage it to think of all different facets and ask me questions about everything. I’m looking to go beyond what occurs to me initially, so I purposefully ask for something thorough that covers every angle.

Phase 2: Critical Analysis and Streamlining

Here’s where I learned that more isn’t always better. I take that maximalist prompt and run it through a critical analysis, asking: “Will this achieve my goals? What can be condensed, eliminated, or optimized without reducing effectiveness?”

I discovered this need for a second step through experience. At first, those exhaustive prompts seemed to cover everything. But when I actually started using them, I often got bogged down. They were too elaborate, requiring way more work than necessary. It’s very much like having too large of a codebase, where you don’t know what all the parts are doing or how they’re going to interact.

For example, I recently created a prompt to help me learn Pixelmator Pro while working on cover art for an upcoming music release. The exhaustive version was thorough and felt like a taskmaster (very step-by-step and tutorial-like). When I ran it through the critical lens, I realized what I actually wanted was hands-on practice and enjoyment. I wanted something exploratory and fun, more coach-oriented than tutor-oriented.

The streamlined version eliminated unnecessary intricacy while maintaining effectiveness. The AI even suggested removing some Risograph brush techniques I’d mentioned being interested in, because incorporating them would have derailed the core learning experience. When I saw that analysis laid out, I realized it was absolutely right. That would have been too much.

Elaborate prompts often become counterproductive because more instructions create more opportunities for conflicting or contradictory elements. I learned this the hard way when I asked an AI to be “encouraging” while also telling it to “challenge me and not always agree.” That made responses too sycophantic, which drives me crazy.

Multi-Lens Analysis

So: streamlined, focused prompts consistently produce better outcomes than elaborate ones. The success of this approach led me to experiment with breaking down other AI interactions into distinct analytical stages, which has become what I think of as multi-lens analysis. I’m now considering how to put all of my chats through various lenses, intentionally approaching the same content or problem from different angles in deliberate stages.

For example, the core lenses I used above are:

  • Generative Lens: Encouraging and idea-focused, aimed at creation and exploring possibilities
  • Critical Lens: Supportive but analytical, looking for improvements, gaps, and optimization opportunities

Another example might start with a pattern lens, followed by a technical lens. I used this sequence when I took the data from all of my posts to see what insights I might find. My first pass was about identifying themes, patterns, and opportunities; my second was about looking at things like length, structure, readability, and tags. Previously, I would’ve put it all into one big prompt and asked it to do everything, instead of taking a more iterative approach. I think this works better.

I also realized that while I often completely disagree with the AI’s analysis, that disagreement is valuable because it forces me to examine the reasons behind my decisions. When I was using an AI as a nutrition coach, the critical analysis kept suggesting dramatic dietary changes and treating my situation like a crisis that needed immediate intervention. When I pushed back, I realized what I actually wanted were sustainable lifestyle tweaks to improve my eating habits, not a complete overhaul. The AI was being much more prescriptive than I needed and assuming I was in worse shape than I actually was. That disagreement helped me clarify that I was looking for gradual improvements, not transformation.

Even when I reject the AI’s suggestions, the process surfaces considerations I hadn’t thought about and makes me more mindful about my choices. The key is being intentional about which lens you’re using at any given moment, rather than trying to get everything from a single interaction.

The Broader Impact

These three approaches have created a major shift in how I think about AI interaction. I almost never stop at the first result anymore. I think a lot of people ask a question, get an answer, and either decide it’s not very good, good enough, or great, then move forward based on that single interaction. That’s how I did it for quite a while.

Working effectively with AI really needs to be an iterative, multi-step process where you focus on one goal at a time in a purposeful sequence. This can be true for research, analysis, creative work … pretty much everything other than casual chats.

It makes sense, too. If you can get a prompt from 1,500 words down to 750 words or less using this process, the AI is going to have a better time with it. The more succinct and clear your instructions, the more effective the interaction. There’s definitely a point of diminishing returns when it comes to context and information when working with LLMs.

What This Changes

This evolution has transformed the nature of my thought partnership with these tools. Voice transcription makes brainstorming and problem-solving feel natural and conversational. Multi-phase prompt development ensures I’m getting focused, effective results rather than overwhelming elaboration. Multi-lens analysis gives me a focused sequence of interactions that allow me to improve outcomes each step of the way.

The combination creates a workflow that feels much more collaborative and less transactional. I’m not just asking questions and getting answers. I’m thinking through problems with a weird, alien-brained partner that can offer different approaches and help me organize intricate ideas while maintaining full human control and ownership of the work.

What I’m most curious about is where this new thinking might lead. What other “lenses” might be useful beyond what I’ve already tried? For now, these three shifts have genuinely changed how I approach complex thinking and creative work, and made the whole process a lot more enjoyable.