Hi. You haven’t heard from me in a while. I’m Carl Tashian. I write about technology, leadership, and the engineer's journey. There’s a good chance you subscribed via my blog.
Living in San Francisco during the rise of AI over the past 3 years, I have secretly and sometimes loudly judged it as boring and soulless. Friends would show me their clever and creative AI side projects, and these things just didn’t light me up. And I would wonder, “Is this where I get off the high-speed train of tech, move to the mountains, and become a shepherd?”
But I have never seen so much hype for a new technology. Such a wide range, too, of doomsayers and optimist milking the hype cycle, competing for whose sci-fi narrative is the most frightening or utopian. And everyone is tapped into it. My uncle forwarded a Substack post “Four Bad AI Futures Take Root” by Brian Merchant. My friend Chris sent the harrowing AI 2027.
I’m an optimistic realist. I try to balance the two. And for me, the potential for deepening social inequity and labor market shocks are my current biggest concerns with AI.
Oh, and the lies. I find it very frustrating that AI assistants lie so thoroughly and confidently. It lies in a way that’s an honest reflection of how people can confidently lie to each other. My hope is that the widespread lies will encourage people to think critically. My fear is that they won’t.
The classic technology history book More Work For Mother gives examples of how new technology is almost always a mixed bag. So often, innovation brings pain disguised as relief, with many winners and losers, and an impact on the world that is never evenly distributed.
This year, I’ve finally turned a corner in my relationship with AI. Or, rather, my several budding relationships with several AIs. It all still feels so new.
Let’s start with programming projects. At first I had ChatGPT write sophisticated Bash one-liners or short Python scripts. It saved me a bit of time, but it didn’t feel like a breakthrough to me.
And then, one day, I needed a PowerShell script for a project. And I don’t know PowerShell at all. So, I wrote a prompt. ChatGPT spit out some broken half-baked code. And even though I don’t know any PowerShell, I’ve been a programmer long enough that I could spot all the bugs and issues right away. So I gave feedback, and we kept working. And eventually, we had something.
ChatGPT had not only saved me time, it had extended my reach as a programmer. And it turned an unpleasant task into a fun project where I felt like I had a collaborator. Normally, especially with an unfamiliar programming language, it’s a lot of work to get up to speed on the language and I have to go through hours that feel like being lost in the woods. This project didn’t feel like that at all. It felt good, flowy and rewarding in the way that programming can feel when I know the language very well.
After these initial successes, I was motivated to try a bigger project: an in-browser synthesizer. As a synth nerd, I came across the Adventure Kid Waveforms (AKWF) one day. AKWF is a giant collection of single-cycle waveforms pulled from many sources. Each file is a WAV with 256 samples. But, there’s no way to listen to any of the waveforms without loading them into a synthesizer.
So, in December I built a 2-axis morphing waveform player in an afternoon with Claude Sonnet 3.5. The experience was very fun, though it involved a lot of copying and pasting (I wasn’t using a coding agent yet).
The program worked, and I learned a bit about AudioWorklets, More importantly, I felt engaged and I had fun.
I had tried building this project just a few months prior, with ChatGPT, and I couldn’t get working code out of the AI. In just a few months, the code quality had dramatically improved. This is what people mean when they say that AI is progressing very quickly.
Last night I installed Claude Code for the first time. Claude Code is Anthropic’s CLI tool for building complete apps—their answer to Cursor, in a way. It uses the Opus 4 model released in May. It’s very good. I wrote an app for my mom, to help her build a catalog of dad’s vinyl records, using AI to identify them from photos. It took no time to have a working JavaScript app, accessing two APIs, with a full React interface and SQLite database. It’s a project I never would have built before. And once again, I had fun! I look forward to the next iteration of it.
Programming feels fresh again. And for someone who has been doing it for decades, that’s a really nice feeling to have.
On this topic of vibe coding, I recommend Fred Benenson’s recent article The Perverse Incentives of Vibe Coding.
Now let’s talk about making AI assistants more useful.
In a recent interview, Sam Altman of OpenAI suggested that people in their 30s and up are using ChatGPT as a replacement for Google, starting from scratch for every chat, asking simple questions that draw upon the model’s knowledge, and getting limited value from it. Easier than wading through Google, but not really tapping into the power of the technology.
Meanwhile, he said, folks in college right now are using ChatGPT as an operating system for life, crafting long, reusable prompts and using them to make major life decisions.
I’ve been through the “replacement for Google” phase, but I’m beginning to level up to the Gen Z wisdom here.
But, how to develop a good reusable prompt?
Well, thankfully, Claude is an AI expert. So, I asked it to play the role of AI expert and help me bootstrap my own AI OS. It started asking questions. About me, about how I want to use AI, my interests, my values, and so on. Eventually this led to me sitting in the bathtub and talking to Claude using its dictation mode. After a few minutes of me rambling into the microphone, it gave me a draft of a master prompt with sections like “my strengths”, “key values”, and “growth areas.” It wasn’t perfect. But what surprised me was how quickly Claude synthesized and organized my long rambling messages into a solid, coherent first draft.
This got the flywheel going for me, and what started as a brain dump has turned into an illuminating conversation thread. I was learning about myself from its reflections.
It excels at distilling and synthesizing ideas. Going up and down levels of abstraction as needed. I’m using it to set goals for myself that help me bring some consciousness to areas where I am not operating in alignment with my higher self. I’m generally not using it as a therapist, though.
My reusable personal prompt is about 1000 words right now and it lives in Claude’s Profile preferences. Because building a reusable prompt is a bootstrapping problem, iteration is key: Revise the prompt, test the prompt, repeat!
And this is how Anthropic gets me to sit down and give them all the details of my life, my personality, my desires and hopes and dreams and fears and weaknesses. And I’m paying them for the privilege.
I like making computers do things they weren’t designed for. With Claude, that has meant making absurd requests. Today I asked it to give me a detailed plan for organizing a bag of USB cables, and it returned a 4 phase systematic implementation plan with an auditing process, length and weight optimization strategies, suggested metrics, and how to use a memory palace to develop a mental map of my cable bag.
I laughed out loud. What a terrible use of my time. I can see that Claude will race further down any rabbit hole than I’m willing to go. And it is not going to step back and say, “Hey, hold up, why are you suddenly so uptight about cable management?” or “Have you had a good hug today?”
Unless I ask. “Should I pursue this project right now?”
It really took me to task. And as it issued forth a very strong no, it asked: “Are you using this project to avoid deeper work on your primary goals?”
Yes. Yes I am.
This is great! Although I'm new to Chat (whatever it is), I really enjoyed your article here. Thanks! Mom
Haha! I didn’t use Claude to write this either, but I too need to procrastinate and have cables to organize.