Positioning – what I might call defining your startup in relation to the problem you solve and environment around you – is one of the most important exercises in Product work as I see it.

How you describe a product should be directly led by what you learn from who it's for – and should directly influence how and what you build.

Put simply: positioning is the meaning at the core of your product.

I can't help but constantly reverse engineer how I think teams ended up with the positioning they have – and what feels right, versus what I might do instead.

I thought I'd write some of this down about the latest Y Combinator batch as an exercise to practice a bit in public and to force me to articulate why I think certain things work or don't, and how they might be done another way.

Over the past few years, whenever I write something, I consider it primarily a kind of "alt take". When bands used to play their songs live for recordings, they would often explore the same premise but keep more than one "take" of the song. I like this concept.

That's a long way of saying: this is just my take. Who cares what I think, it's not like the only determinant of success is how much Max likes your YC batch page. Your mileage may vary.

Now, on with the show, in no particular order:

TokenOwl: AI TurboTax for the next generation of crypto

• Why is it AI? What does that have to do with how the product works? If it's TurboTax, it being "AI TurboTax" doesn't immediately mean anything, and certainly doesn't mean anything more or better.
• What is the next generation of crypto? The next gen of crypto tokens or the next gen of crypto investors?

TL;DR: TokenOwl modernizes tax reporting for a new generation of crypto traders. DeFi users and high-volume memecoin flippers can now use an AI assistant to calculate taxes with greater accuracy and find trends in their trade history.

• Ah OK so it's an LLM assistant for high volume crypto people to manage their tax.

We’re building TokenOwl to save traders time and money. 💰

• More or less every piece of software ever written exists to save time and/or money. It's basically a phrase you can't use meaningfully in to position yourself because when something is part of everyone's core sell, it doesn't add anything to say it's part of yours.

Current crypto tax software solutions work fine for traditional buy-and-hold crypto users with dozens or hundreds of transactions per year, but they aren’t built for DeFi users and high-volume memecoin traders. They tend to mislabel these transactions which can lead to capital gains calculations being incorrect. The only way to fix such mistakes is for the user to spend hours–sometimes even days with enough transactions–troubleshooting their tax report. They also cost too much for users with thousands of transactions.
As the IRS increases scrutiny on digital assets, reporting crypto capital gains and income accurately will be crucial.

This could be shorter and clearer:

Because DeFi and memecoin enthusiasts trade in such volume, existing crypto tax software often mislabels transactions, resulting in capital gains mistakes unless they spend hours troubleshooting their tax reports.

LLMs are really good at processing large amounts of nuanced data (like complex DeFi transactions). We are building a solution that leverages LLMs to reduce breakage when generating tax reports and minimize the time spent on troubleshooting/understanding transaction logs. Users can also query their logs directly with natural language, to unlock new possibilities in analysis and review. This makes for a smooth experience for both experienced traders and occasional investors who need help parsing their portfolios.

Ah. Now we have a problem – because it seems like all this product does is use LLMs to get a better understanding of high volumes of crypto trades.

I'm bored of crypto anyway, let's move on.

Keywords AI - Unified DevOps platform to build AI applications

Dead simple way to deploy & monitor AI apps with 2 lines of code.

• Don't use DEAD in the first line of your product pitch. When we're in understanding mode, our cognitive process is raring to go and soak every single word dry for meaning. This is not a word it loves. Unless you are going to introduce it in just the right place, for a kneejerk shock impact, you need to be really careful with it.
• Furthermore, isn't it better if you remove it: "Deploy & monitor AI apps with 2 lines of code."

‼️ The Problem

LLM developers face many challenges that hinder the ability to quickly build and scale high-quality products that people love:

- Time-consuming infrastructure setup and maintenance for deploying and scaling LLM applications

- Difficulty collaborating on prompt engineering and evaluating models for specific use cases

- Challenges monitoring and ensuring consistent performance in production for optimal user experience

- Tedious process of creating datasets for fine-tuning and iteratively improving output quality

• Never have more than three bullet points in a list. If you need more than three, you aren't making choices, or you aren't consolidating ideas into bigger themes. Very quickly, this might be something like:

1. Setup, maintenance + monitoring for LLM apps is time consuming
2. Prompt engineering and model evaluation struggles for specific use cases
3. Creating datasets for improving output quality

• But even so, we are quickly running into the limitations of what I think happened here: they worked back and wrote their problem statements from what they were planning to build anyway.

Keywords AI is built for developers. It has every feature needed to build, deploy, and scale LLM applications:

• How is it built for developers? What exactly have you done to make that true vs anyone else who might build this for developers? What does it mean to be built for developers? In fact, your key feature of having only two lines of code makes being a developer even less important to get this working.
• Saying you ever have every anything for anyone is nightmare positioning. It means again you haven't made any choices. There's no way you have ever feature for every user, especially as you just graduated YC. And it suggests you don't know what's really important for developers, otherwise you would build that first and focus on it first. Which suggests you don't really know what the deep and direct problems are that they face in this area – so you've just gone broad.

ProSights: AI-Powered Analyst for Investment Firms

Making sure you never miss a good deal

• Nice – everything in the message so far feels like very low cognitive load, I was able to absorb it quickly, I could tell someone else what this is about, and it doesn't matter the sub-head is pretty low value, as I've already got momentum and I know we're about to get the proper pitch.

ProSights is building an AI-powered analyst for investment firms. We use LLMs to harness all your data sources so you can seamlessly track portfolio companies, automate tedious daily tasks, and receive alerts about new companies taking off.

• "Harness" is actually a better word and metaphor to establish here than it may first seem. The idea of literally putting LLMs as a harness over a stable of data sources feels like a sleigh full of reindeer.
• Three obvious big feature threads even up here, although I think they naturally simplify to two: 1. track companies (portfolio and new), 2. automate tedious tasks. The only thing is, these feel like quite a generic combo I might achieve with a lot of tools – is there nothing else investment analysts do, or crucial in the way they do it that we could focus on?

Many investment opportunities slip through the cracks due to poor coverage.

• Having worked inside VCs, I think this is indisputably the thing that keeps them up at night and what having analysts tries to help them resolve. Nice.

Investment firms sit on a treasure trove of data (investor presentations, company financials, external data sources) that’s often hidden in old email threads and obscure folders. As a result, it’s challenging to stay up-to-date and make fully informed investment decisions.

• I think there may be scope to relate the first point and the second more closely as part of the same puzzle – some money left on the table here. Let's try:

Investors can only win deals they are aware of, and understand. That means: having a comprehensive view of the market, and learning from their data and experiences to make the best possible decisions.

Both of these are made more difficult by huge and growing volumes of unstructured data. Both can now be solved more effectively than ever, thanks to LLMs.

- interactive dashboards,
- find important documents from your team’s email and files,
- Smart Actions Dashboard makes sure you never miss an assignment, 
- AI chat box answers any question,
- alert you when a company is taking off

• These summarise the bold text in the solution section. Again, we are drifting into more than three key product pillars, although to be fair some are lumped into the same paragraph. However, it's starting to feel like this is kind of an old-fashioned SaaS tool. You can search for docs, dashboards tell you how things you are tracking passively are doing and it alerts you when something kicks off. What's special? Where's the magic? I think you'd have to sell me with numbers to show that this does X or Y 10x better, or that how I pay for it is smarter – or that I can't hire some devs to build it for me just like these guys did the first time round at Permira.

If you’re an investment firm that would find this useful, please join the waitlist

• We're getting into broader messaging stuff here but I thought it's worth noting how waitlists have become more and more common. If you're aiming for high esteem/ high prestige audiences, this is a good way to make them feel completely worthless. Unless you are truly dying for capacity, can you reframe this that they might actually get a response quickly and from a human being? Exclusivity can provide prestige – but waitlists rarely feel like that in B2B now.

K-Scale Labs: Open-source humanoid robots for everyone

Meet Stompy, the world's first open-source general-purpose humanoid robot

• Rules are made to be broken. Usually you don't want to position for everyone – in this case, it's great.

Topo: AI sales agent custom trained for each company to reach the best leads

• Hmmmm. Isn't the whole problem with cold outreach today that you end up with an inbox full of bullshit from generic incoming and you mark it all as spam. Too much noise. Reaching the right people doesn't seem to be the problem anymore, breaking through the noise does.

B2B lead generation is broken. It's difficult to generate outbound pipeline as we did in the past years: mass mailing over the TAM based on demographics criteria. It's especially very human-intensive for a founder, a full cycle sales team, or even an SDR dedicated to the task. Finally, the cost of implementing a high-quality outbound strategy has completely exploded.

• "X is broken" was cool in 2009, but even then it has a central flaw: you are telling me instead of showing me. Delete it and show me how it's broken.
• Why is it difficult to generate outbound pipeline now vs previous years? Why has the cost exploded? Go deeper.

AI allows us to automate these tasks by doing like a top performer.

What makes a top performer?

They know the product, the industry, the pain points, and the company's value proposition perfectly.

They will find the right leads, ready to buy, based on buying signals in the industry-specific channels or on the company's website visitors.

For example, the buying signals for a dev tool will likely happen more on GitHub, whereas for an HRIS, it will happen on LinkedIn or job boards.

That's what we do: Create the smartest AI sales agent on the market - custom-trained for each company!

• Argh. It's like the problem we have set up is: "outbound sales is harder than ever" and our solution is "make a good sales AI!"
• I feel like we need to totally start again. What's the purpose of outbound sales? Why has it got so much harder recently, what's the trend? Why is AI in particular the fit for that particular recent change? Why is the way YOU can use AI going to be a better answer than any other GPT efforts? I'm just not convinced by anything here because there's no insight in the setup or solution.

Reprompt AI: Helping AI apps get to production quality 10x faster

AI teams use Reprompt to evaluate and improve their apps without code

• I LIKE IT. I LIKE WHAT I'M SEEING. It feels simple, it clearly states who it's for. "evaluate and improve apps without code" is so easy to absorb, and makes me want to learn how. "10x faster" is great (as long as they can deliver...)

90% of Generative AI pilots never make it to production

• Yes. Uses data, sets the scene, feels obviously frustrating just to hear, even if you have never been involved in a project like this. I'll even forgive the 4 bullet points that come after.

Getting to production takes more than great prompt engineering

• Great principle, said with confidence.

Reprompt helps AI teams fix “last mile” AI issues so they can ship faster

• Aaaah, fantastic. Repetition and adding detail is going to make sure I leave with a consistent idea of what you do.

Here’s what you can do with Reprompt:

1. Trace your AI calls across chat, RAG, and function calls to make debugging easy

2. Automatically track and highlight hallucinations and file them as bugs

3. Write custom prompt overrides to handle the edge cases without changing your main prompt or pushing code

Beautifully clear. A list of three! A caveat in the third point expressed simply and that increases the specificity. Yes.

OK, that's 2500 words or so, enough for one day.