Subtle Software: Start here

Welcome to Subtle Software - a blog about building a one-person company using LLMs. Here are some posts to read first.

Notes on pricing software

Lessons learned about designing around price, why B2B is easier, and how to price under the corporate approval bar for easier sales.

Adventures in automation

Automation is especially important for solo businesses since it's just you. Here are some automations I set up while building a data subscription product, including crawl jobs and an autoblog. I also outline my ideal auto-marketing product.

Asking people for money

A guide for introverts.

Strengths and weaknesses of subscription data products

I'm building a subscription data product and exploring the strengths and weaknesses of this business model. Subscription data products like Nielsen, App Annie, and Bloomberg have key advantages: they offer a cornered resource, scale efficiently, and accrue value over time. But they also face challenges from easier scraping tools and limited TAM...

Post-LLM sales and marketing

Software is becoming much easier to create, with many second-order effects. As Matt Mazeo highlighted on My First Million, AI dramatically lowers the cost of work, enabling businesses to sell finished products rather than promises. Post-LLM marketing should deliver product value without requiring action from the recipient...

Subtle Software v0.2: Reflections, posting cadence, time to build, and services

After taking time to reflect on Subtle's journey, I'm adjusting the posting frequency to 2-3 times per week and focusing more on original products. I'm also introducing Subtle Services, offering product management coaching and software development services...

2k tokens per second

Google's demo of a coding agent producing 2,000 tokens per second has significant implications for the user experience around generative app experiences. This breakthrough could transform how we interact with text-to-product apps like Replit and Lovable, making the experience more magical with instant feedback...

There is a massive dislocation in American knowledge work coming

As AI becomes capable of producing high-quality work at minimal cost, American knowledge workers face a fundamental economic challenge. Microsoft's recent layoffs signal the start of a major shift in how companies value human expertise, with language models offering a compelling alternative to expensive human labor...

Managing agents ≠ managing people

As we enter the era of cybernetic teammates, traditional management skills like EQ, charisma, and executive presence become less relevant. The key skills for managing AI agents are different: giving specific instructions, setting clear expectations, providing examples, and evaluating work...

Replit could win the vibe coding market

Replit introduces new security features for vibe coding, positioning itself between Lovable and Cursor in the code generation space. Their web-based IDE and embedded cloud services could make them the ultimate winner in the AI-powered development tools market...

Every (the company) knows what's up

Every has built a 7-figure AI consulting business by combining good writing with hands-on product development. They earn credibility by shipping their own products and learning firsthand how these tools work in practice, which gives them unique insights into helping companies adopt AI effectively...

Klarna's semi-reversal and Amara's Law

Klarna's recent shift in their AI implementation strategy reveals important lessons about enterprise AI adoption. After initially replacing human agents with AI, they're now taking a more nuanced approach, highlighting the gap between AI hype and practical implementation...

An interpretation of OpenAI buying Windsurf

OpenAI's $3B acquisition of Windsurf reflects two strategic imperatives: extending their product capabilities to capture the "workspace of the future" opportunity and expanding their enterprise beachhead. This move could significantly impact the competitive landscape for AI-powered development tools...

Jony Ive on the spirituality of careful design

Jony Ive shares a profound perspective on design: how small details, crafted with care and love, can create meaningful connections with millions of users. In a recent interview with Patrick Collison, he reveals how thoughtful design is a way of expressing gratitude to humanity...

Landing pages are dead

The title is slightly provocative, but I wanted to collect a few observations about the psychology of discovering and adopting new products in this new age of software development. If I think about the tools I've personally adopted as daily workhorses – ChatGPT, Cursor, Devin – I was not convinced to try them by a landing page...

The emergence of the "software creator"

As new tools allow individuals to contribute in domains outside their core competency, we'll see the traditional product trio roles merge into something akin to "Software Creator." This parallels the creator economy, where new publishing tools allowed independent media creators to compete with incumbent media companies...

If we're all gonna eat, someone has to sell

Ken Griffin, founder of Citadel, recently gave a talk in which he emphasizes that entrepreneurs must get comfortable with selling - to candidates, vendors, counterparties, and customers...

Software agency prices are about to collapse

Tools like Cursor and Lovable are making it way easier to build software. Software agencies will not be able to justify their previous prices and timelines in this new world. Previously, agencies could quote low/mid six figures for 3-6 months of work, but now a solo developer can build the same product in weeks...

The perverse incentive behind Lovable's growth

Lovable is reportedly the fastest growing startup in the world. After trying it out, I noticed their extremely aggressive paywall - just 5 prompts per day on the free tier. This creates a perverse incentive: if Lovable became excellent at one-shotting users' app concepts, fewer people would hit the usage limits, reducing the growth of their paid tier...

Agency, taste, and new creative tools

"Agency" and "taste" are buzzwords these days, but their popularity hints at how the creative process is changing across many domains. While AI tools are helping some people create slop, the best creative tools will enhance creative expression by giving creators high-fidelity representations of end-user experiences throughout the creative process...

Cursor for X

"Cursor for X" is a popular framing for new products. Just as Cursor is changing the nature of software engineering, new AI-native software will change the nature of other job functions and indsutries...

Windsurf is an enterprise company

Windsurf CEO Varun Mohan recently appeared on the Lenny podcast and shared insights about their enterprise focus. Despite consumer word of mouth, Windsurf is primarily an enterprise company, with 80 of their 150 employees in sales. They've built specialized infrastructure for large codebases and work with major clients like Dell and JPMorgan Chase...

Creating and maintaining a prd.md file

I've previously written about the utility of having a prd.md file in your project repo. It solves the problem of new chats with code agents starting with zero context. Here are some tips on creating and utilizing this file effectively...

Notes on using o3

OpenAI launched their o3 model last week, and Tyler Cowen called it AGI. While the AGI debate is semantic, it's clear we're in the age of AGI. I share some practical observations about using o3, including its verbose nature, technical depth, and how it differs from GPT-4 in terms of engagement and dialogue style...

DistillJS

I built DistillJS to recreate specific LLM behaviors in Javascript. It works by challenging o3 to write a minified JS function that yields results broadly comparable to gpt-4o-mini's on a specific task. I just launched it (in alpha) so try it out and email feedback...

Sunsetting stringme.dev

I built stringme.dev a few weeks ago to generate a plain text version of any webpage in a format optimized for LLMs. Firecrawl recently introduced their own version of basically the same tool: llmstxt.new. While there's been some traction on stringme.dev, llmstxt.new is a pretty solid solution, so stringme.dev will now redirect to llmstxt.new...

Product-market fit is more transient than ever

Jason Lemkin (SaaStr) and Rory O'Driscoll (Scale) went on 20VC recently and shared their observations about how product-market fit has become more transient than ever. They discuss how SaaS investing has changed, with companies now potentially losing product-market fit in weeks rather than years...

OpenAI, Windsurf, and the future of work

OpenAI is reportedly in talks to buy Windsurf for $3 billion, after previously considering Cursor. Building on Jake Handy's observations about Cursor (and Windsurf's) potential to expand to non-developer work, I'd argue OpenAI is bidding on Windsurf as a play for the future of work...

Tiny teams data roundup

Here's a bunch of new data on small teams with massive revenue numbers, including tinyteams.xyz, IndieHackers' thread on one-person companies, and data on AI unicorns' revenue per employee...

The ChatGPT "projects" feature is underrated

ChatGPT's projects feature launched in December and offers powerful custom instructions for organizing related chats. I've been using it for meal planning and startup ideation, with project-specific instructions and persistent context across conversations...

Problem hunting with language models

I'm reading the book Problem Hunting by Brian Long and exploring how LLMs can make the process easier. Tools like Deep Research and Rally can help validate startup ideas faster and cheaper than traditional methods...

Product management is writing good evals

In a recent talk, OpenAI Chief Product Officer Kevin Weil said writing evals is becoming a core skill for PMs. This reflects both the growing importance of models in software and the changing nature of product management work, where implementation is increasingly outsourced to LLMs...

ChatGPT's new memory feature is all about lock-in

ChatGPT announced a new version of its "memory" feature that will draw from users' entire chat history. This creates significant switching costs, especially in enterprise contexts where chat history contains valuable business data...

Sahil Lavingia's future of work experiments

Sahil Lavingia, the founder of Gumroad and Antiwork, is testing compelling ideas about the future of work, including open source bounties, cybernetic teammates, royalty-based funding, and flexible work arrangements...

Book rec: The Thinking Machine

Highly recommend the new book The Thinking Machine by Stephen Witt. I learned a lot about Jensen, Nvidia, the history of GPUs, and machine learning...

The Shopify AI memo is the future

Shopify CEO Tobi Lütke recently sent a memo to his employees mandating the use of AI tools. The capitalistic logic behind this mandate is irrefutable and will become the norm, especially if economic conditions worsen in the near-term...

An exit strategy for solo founders

Josh Mohrer was on the Generative Now podcast recently and floated an exit strategy for solo founders. If you successfully bootstrap to profitability, it might be best to sell to a company that wants to go multi-product, which is happening earlier and earlier...

Using Clay for lead gen

We're doing lead gen for Flow State sponsorships, and Clay is really good. They raised a "pre-emptive Series B expansion" at a $1.25B valuation earlier this year...

ChatGPT should add user-to-user messaging

Zach Weinberg, who founded Flatiron Health and Curie Bio, floated an interesting idea for OpenAI in a podcast today: adding user-to-user messaging to ChatGPT as a form of non-monetary lock-in...

Flow State's quarterly R&D post

Our friends over at Flow State put out their quarterly R&D post for founding members. Highlights include surpassing 40,000 subscribers, revenue growth through their sponsor booking page, and plans for a potential indie streaming service...

What is "good" AI revenue?

Recent reports of questionable revenue practices at highly capitalized AI startups raise important questions about what constitutes sustainable growth in the AI era...

Gemini 2.5 excels at refactoring code and higher-level projects

On Tuesday, Google announced Gemini 2.5, a reasoning model that is optimized for coding and has a 1 million token context window.

Cursor made it available today, so we decided to give it a shot.

We saw that for another Cursor user, Gemini 2.5 was able to solve a problem that Sonnet had gotten stuck on...

Microsoft's security agents

I think Microsoft has the right idea with their security agents. This could help the urgent "vibe code exploits" problem...

Vibe code exploitation

Fireship tackled vibe coding in a recent video. There's an urgent need for a turnkey cybersecurity agent in the age of AI-generated code...

How to get startup ideas in 2025

Paul Graham has a classic essay about how to get startup ideas.

"The way to get startup ideas is not to try to think of startup ideas. It's to look for problems, preferably problems you have yourself..."

prd.md

A common pain point working with code generation tools like Cursor is that with every new chat, the agent needs to scan the code base to get context. The agent acts as though it's never seen this codebase before, even if it wrote every line of it.

The best way to mitigate this is to include a product requirements doc (PRD) as a markdown file in the repo...

The cybernetic teammate

One of our observations at Subtle is that, enabled by LLMs, "individuals will increasingly outperform teams." A new paper from HBS, Wharton, and Proctor & Gamble tested this idea in practice.

Researchers worked with 776 employees at P&G, randomly assigning employees to work with or without AI and work alone or with a team...

The vibe coded landing page

There is a telltale sign of a vibe-coded landing page: header (product name, call to action), h1 greeting, h3 description, another call to action. They all look the same.

Check out some examples of vibe-coded landing pages...

Is Cursor a good business?

Sam Lessin offered an interesting, skeptical take on Cursor and similar LLM wrappers recently: "I like [Cursor]. I use it as a product. I don't see it as a good business.... There's no lock-in.... I think the whole point of AI is the switching costs for all this stuff goes to zero."

Even though Cursor may be the fastest startup to reach $100M in ARR, it does face some existential risks...

Vibe code responsibly

Vibe coding is not without risks and may lead to lost work and security vulnerabilities. Here are some pointers for how to vibe code responsibly.

That said, there's also a new security opportunity in the world of superabundant software...

stringme.dev changelog – 3/17/25

Today Stripe added the ability to copy the text of their docs or view the page as markdown. This is an example of a company proactively making the contents of their webpages more readily available for LLMs.

However, this is one of over 50 billion webpages. There's a tragedy of the commons problem: most webpages won't make /llms.txt or equivalent available anytime soon. In the meantime, that's what stringme.dev is for...

How it's made: stringme.dev

The main lesson from building stringme.dev was to identify the languages, frameworks, and services you'll need up front when vibe coding.

We ran into problems when we told Cursor mid-project that we wanted to use Tailwind, deploy via Vercel, or switch from TypeScript to JavaScript.

Here's how we vibe-coded stringme.dev and what we learned along the way...

stringme.dev updates

Spent today iterating on stringme.dev.

Added quick URL buttons for example websites, improved copy across the site, and implemented Mozilla's readability library to strip away extraneous filler text from webpages.

Also improved text optimization for LLMs while stubbornly refusing to use an LLM for this process itself.

Check out all the updates and improvements...

stringme.dev

Here's something new from us: stringme.dev.

This service takes a URL and generates a text string. Sounds simple. Why does it matter?

I got the idea from an Andrej Karpathy tweet about needing a service that takes a URL and returns clean text content for LLMs.

Check out the full story behind this new tool...

Simon Willison on using LLMs to write code

Simon, who created the popular Python framework Django, published a long blog post today about using LLMs in his coding workflow.

He's been using LLMs for about two years now and has some interesting insights:

Once I've completed the initial research I change modes dramatically. For production code my LLM usage is much more authoritarian: I treat it like a digital intern, hired to type code for me based on my detailed instructions.

Learn about his approach to using LLMs for research, coding, and more...

Some interesting new products

Manus is another impressive Chinese LLM product, following DeepSeek. It's similar to OpenAI's Operator product, in that it's an agent that can be assigned a complex task and then autonomously, iteratively complete it.

Sesame is an extremely impressive voice demo.

Check out more interesting products including seven39 and Mastra...

The Collisons are great writers

Stripe just published their annual letter.

The main point I wanted to discuss is their defense of LLM wrappers. But I couldn't help but admire their writing abilities:

For too long the crypto economy was an isolated atoll, with vibrant native customs but few exports to the rest of the world.

Clearly the Collisons write without LLMs. They raw dog the page like Mark Twain...

A subtle approach to newsletter sponsorships

A little newsletter called flowstate.fm asked us for help with their sponsorship business.

They wanted to make a website where people could book weekly sponsorships directly. They also wanted some help with outbound lead generation.

The result is sponsor.flowstate.fm which we built in just a few hours. Here's how we did it...

The solo team reward function

When you're working for a company, your compensation is typically a salary and maybe a bonus and/or stock options, with performance reviews determining your success. But in one person teams, the only priority is customer service – compensation is exactly company revenue, determined entirely by customers rather than managers...