We’re hungry for newsletter feedback! Please do this 2 minute survey on the newsletter, so we can improve and give you more of what you want.

Onto the newsletter!

AI hype versus denial

The world is breaking into a AI hypers and AI deniers. The hypers think AI is going to do everything, is better at everything–except building data centers, which will be all humans are for–and can’t stop telling everyone about it on socials.

The deniers are hunkering down in their caves, trying to figure out if rubbing sticks together is better than using flint to start a campfire–without asking ChatGPT for the answer.

I exaggerate. Or do I?

It reminds me of an argument I had with my parents in 2005–was it faster to Google something or use the Yellow Pages? Back then, in their house, Googling actually required going to a room with a computer in it–so we raced. Google won, but it was close.

That is obviously ridiculous today.

Almost as ridiculous as most peoples’ expectations of AI coding tools still are.

I’ve been writing a lot here and on LinkedIn about some cutting edge AI GTM stuff–like MCPs. But AI coding tools are still out there, they’re still better than ever (maybe 100x better than even 3-6 months ago), and I still think most people are underestimating them.

Vibecoding can do big things

Vibecoding, or coding with AI tools, had a moment. The article I wrote with Kyle Poyar was his most popular of all of last year–500 LinkedIn reactions, 100+ comments. We saw over 150 signups and I didn’t even talk about Skyp (it wasn’t even really live yet–oops!).

But vibecoding seems to have lost the limelight, at least in GTM circles. Instead, my LinkedIn feed is full of lists of Claude Skills. These are recipes, essentially, that tell Claude what to do–what specific tools to use, what to ask a user, etc. You could build a skill to refine your ICP or your pricing, for example. And you could–like Kyle Poyar did in that post–share it with the world. It’s pretty cool, actually.

But I’m not ready to let go of software, specifically things you can make with Claude and other AI coding tools (Bolt, Lovable, Replit, etc.). So I thought I’d share an example, and how I built it, to give you an idea of what’s possible.

The hard part of building software with AI

The coding tools have become so good that the hardest part now is knowing what you want to build. Whereas getting Bolt to build a decent website a year ago was an exercise in frustration beyond the first couple of pages, today you can “one-shot” huge and complex tasks.

“One shot” means basically put in one prompt, have the AI plan it, and then build it. Other people may use this phrase differently than I do, so to define terms: to me this is not a one-sentence prompt like “Build me a CRM” but rather an extended back and forth, that results in an actual “plan” (a markdown or “.md” file) that Claude Code or other AI coding tools will refer to and work from.

The key to the “one-shot” is that if you build it right the first time, it is infinitely easier than building something that is 60% right and iterating. Because this is actual code–if you change your database structure, you have to actually migrate your database. Even if AI does it for you, it’s still complex.

So most of the effort–90%, in my experience–now goes into figuring out exactly what you want to build. If you want to build a CRM, what are you building it for? What jobs to be done does it need to do? What are you omitting that Claude might expect a CRM to have? This is now the hard part.

The other catch

If you are building Kleenex software–something you’ll use once and throw away–the only hard part is getting it to work the one time. But if you are building something you plan on reusing–like a CRM–you will find that maintenance becomes important.

There are projects I’ve built to solve a one-off problem, realized that I had that problem frequently, and then tried to use them on an ongoing basis. For example I have a project that manages Google Ads (with, admittedly, mixed results). Its utility is mostly two things. First, it has scripts to upload, download, and transform files in the formats Google expects. Second, it has all of the stuff I’ve tried before, and all of the output of those files, so I don’t have to repeat myself or go get the same data twice.

Because maintenance is a pain. Sometimes it’s worth it. Sometimes it isn’t.

If we take our CRM example, it’s been enough of a distraction that I’m now trying AI CRMs (Lightfield, at the moment).

To me this is only really worth it if the cost savings are massive, if the problem is core to your business, or if you plan on shipping this as a product one day yourself, so getting firsthand experience is valuable.

An example of what you can build

My intention is to show what’s possible. This example is tangentially related to sales–but if you go check it out, you’ll see that it is on par with most consumer apps. I’ll explain my process. You can go to the Chrome store and try it in the same way you could get the Apollo or Hubspot Chrome extensions. Pretty pro, IMO.

I started with a simple problem: I read too much. Between the latest stories in Bloomberg, the occasionally useful LinkedIn post, and all of the newsletters I get (many of which written by people I know) I was losing at least 30-60 min a day to unproductive reading.

I don’t actually need to know every detail of what POTUS did today. I love my friends’ newsletters but I don’t have time to read ten 1,500 word newsletters a week. I just want to be aware in general of what’s going on. Sometimes I’ll dig deeper, but not always.

I built an app for this and posted it to LinkedIn a while back–50 people used it (according to Google) even though it was really difficult to use. You needed a ChatGPT API key, among other problems.

I figured, Claude Code had revolutionized what I did at work. Maybe it was time for v2.0.

Planning

I spent 30-60 minutes designing the app with Claude. I had a bigger ambition this time–though the core functionality remained.

I wanted it to be easy to use. That meant both a server-side component, and using my own LLM API key. Which also meant taking payments (I am not not in the business of subsidizing strangers’ AI usage).

I wanted to ask questions. Sometimes the summary missed something important, or mentioned a big topic but clearly more was covered. Could I chat with the summary? Turns out if you set up a server, you can.

I wanted to be able to share the articles. Frequently the articles or newsletters I come across are relevant to other people–wouldn’t it be great to share a summary? With a server this is possible.

I wanted some growth vector that could be viral. This was perhaps just me being ambitious, but sharing seemed like not enough. What about “top shared” lists? What about getting Google to index the summaries, and get traffic to them?

Once I had a plan, I was ready to review it.

Tom Tunguz’s AI coding hack

To be fair, Tom has a LOT of AI coding hacks, but this is perhaps the easiest of them all. Take Claude’s plan, and give it to another model–Gemini Pro is his (and therefore, my) go-to.

This is super easy in Cursor, if that is where you use Claude Code, because you can just tell Cursor to use Gemini Pro and point it at the plan file.

As a direct result, Claude thinks I am a coding genius. Because I just take Gemini’s feedback and give it to Claude, as if it were my own.

Usually after 2-3 back-and-forth reviews, there’s nothing left to improve.

Read the plan yourself

This is the part where I admit to taking AP CS as a sophomore in high school, and a lot of CS classes at Dartmouth. I kinda sorta know what I’m doing. But with the advances in AI coding, my value add here is under 5%. At this point you could YOLO things and be fine.

That said–I do read the plan. All of the plan. If it’s too long (which it sometimes is) I have Claude summarize the important parts.

Build it–in one shot, if you’re lucky

The build can take an hour, or more, depending on what you’re building. For an ROI calculator or a simple sales tool, it might only take 5 minutes.

Usually something will be not quite right. This is the key learning period, when you thought you knew what you wanted but when you’re in fact presented with that exact thing–you change your mind.

This used to infuriate engineers, but now you can just go change it yourself. I like to think of this a great way to achieve 100% responsibility. You can chose to leave well enough alone, or you can strive for perfection. Both have a time and place.

Just remember that you learn a lot from using software, so if you can get something out there that is good enough–you might feel differently about where to put effort 24 or 48 hours later. Than if you instead spend that 24 or 48 hours tinkering with almost good enough.

The finished product

The summarizer app is called tl;dr. It’s available in the Chrome store, and the free version is super useful–I summarize a lot (100s of articles so far) and share them with friends.

It took about two days (a weekend) to ship it. That included all of the Chrome store stuff that needs doing, setting up a server (I use Railway) and database (I use Supabase), and setting up Stripe (yes, it takes payments, upgrades, downgrades, etc.).

The sharing looks like a professionally made app:

It counts how many people looked, also, so if you’re sharing for professional reasons (e.g., to a sales prospect–a key use case of mine) you can see if they looked.

Stripe was really the most incredible AI win. This was a project we did at my last startup that took two engineers about a month. It took a very smart lawyer to write out the massive if/then tree to handle upgrades and downgrades. Claude nailed it on the first try; humans spent ~0 minutes thinking about it. And, it works.

Don’t give up on AI coding yet

I’m not sure if I wrote this for the AI hypers or the AI deniers, but I hope wherever you are on that spectrum–it was useful.

I think most people–even the AI hypers–are not ambitious enough with what they expect from AI coding. Both in running their companies, and in doing small tasks like this.

A VC mentioned some engineers are spending $1 million per year on AI coding agents. If I built that complete app without coming close to maxing out my $200 Claude Max subscription, what could I have done just that month with $80,000 worth in credits?

Clearly, I have a lot to learn.

For most people, paying $2.99 would be far better than building your own version of this app. But for some, there is some useful tool for your business you could ship using AI coding that would create tremendous value.

Hope this was useful! If you got something out of it, or wanted more–just reply. If you want to talk about Skyp, just reply… we’re here for you and all your AI GTM work.

PS: Want another example? Check out Skyp’s free email grader. Also built by yours truly, that took less than a day. We get several leads from it every day. I would argue–better than any other tool out there, free or paid, if you’re wondering how your sales emails are doing or want to improve them.