Jason Bosco

LLM-generated (slop) PRs are going to fundamentally reshape how open source communities function.

If you’re going to spend your money on tokens for this, please make it a bit more useful for all of us.

To the folks using LLMs to auto-reply to posts:

Might I recommend you prompt your agent to use your tone of voice, and have it say something unique?

All I see is either a rehash of what was said, or a bland follow-up question.

It’s super easy to see through this.

TIL that the Federal Reserve has a status page, and that ACH payments had an incident today.

Never imagined I'd see this statement in an AWS status page

"We used to debate using tabs vs spaces in code we'd type out"

Doing all of this manually used to be such a chore, and offloading it to LLMs feels like I can finally explore our data without thinking about the time it's going to take to pull the reports for it.

Still not one-shotting it by any means, but it speeds up the whole process significantly.

I then ask it to generate some sample rows in a test database and have it trace through the various parts of the query, to ensure each portion does exactly what is expected.

Then I ask it to run an EXPLAIN, look at the query plan and then do performance optimizations so the report runs fast.

I get an idea for a metric I want to explore, describe it to Claude Code, point it at our latest DB schema in version control, and let it do its thing.

One of my recent favorite use-cases for Claude Code is having it write SQL queries for reports.

No more sulking over elaborate JOINs just to pull in that one field that’s N degrees away from the primary table.

My 5-year old and I spent an evening building a little ping pong game using Cursor, just for the two of us to play.

It might not be polished or perfect, but it will be useful. Just to you.

Software is no longer a question of “is this worth building for enough people”. You can build it just for yourself, and your family and your friends.

I’m calling it pico-SaaS.

That app to help you track snack inventory in your office pantry, ask an agent.

That app to help track logistics for just this one event you’re planning. Ask an agent.

If you can describe the problem clearly, you can usually build something useful in an hour.

That 3D visualization you wish existed to use to teach your kid about some concept, you can now ask a coding agent to build it.

That app that can help you save your tried-and-true tweaks to common recipes, ask an agent to build it.

There’s a category of software that should exist, but doesn’t

because it makes no economic sense to spend time on it, even as a side project.

Coding agents are going to massively unlock this niche of software, for everyone.

Feb 2026 way of doing prep work for tax season: using Claude Code

Amen.

The real milestone is to reach a point where you don’t need to raise external capital any more, because you’re happily profitable.

And when you reach that point, the lesser stock you sold to get there, the better.

In our case, we sold ZERO shares before we reached profitability.

LLMs are incredible. And it's completely ok for them to help us build a very good first draft, that we can then chisel and shape into what we want over a longer period of time.

The joy is in the chiseling.

Something about "one-shotting" software products into existence captures our fancy so much

that every hype wave uses that as the yardstick.

You know what they say, Rome wasn't one-shotted in a day.

The dangers of not reviewing LLM generated content. More slop.

(Brand designers, please don't hate me 🙏 ).

My suspicion is that they preemptively chose a shade of green that survives on both web AND print. Their previous brand color would have been brutally bad in print.

In our case, we made the opposite call to sacrifice consistency in favor of picking a color that looks best in each medium.

It took me a few days to get over the difference in colors between our site and billboards. (The details you sweat as a founder!)

Anywho, back to Granola.

I scrambled to a nearby print shop (who happens to be our customer, yaay), and worked with them to pick the closest greenish-yellow color that looked bright enough on vinyl print in CMYK.

So if you notice closely, our SF billboards look more yellowish than green.

So when converted from RGB to CMYK, our nice and bright green ended up looking exactly like the sad green in the image below.

Painful lesson in color scales.

When it came time to design billboards for our SF campaign last year, we happily used the same color and it looked great... on screen!

And then came the *print* proofs of our billboard designs… in the CMYK color scale 🤢

Turns out the green we have in the RGB scale does not exist in CMYK.

I have a plausible theory on why Granola chose a seemingly worse shade of green in their new redesigns.

We went through the same pain with Typesense.

We use a nice and bright shade of green on our site.

Ok this is getting out of hand, for better or worse.

This is a site where an AGENT can hire a HUMAN

To do things on its behalf in the real world, that it can’t do by itself.

This is like a new episode of Black Mirror unfolding every day

Notice how there are two separate CTAs for humans & agents

So in a future where agents write majority of the code, I think I'll still enjoy writing code by hand... just because.

Yes yes, I know software engineers are hired to solve problems and not just write code.

But there's still a good part of me that actually enjoys solving problems *by writing code*. I enjoy both.

Because if I only enjoyed solving problems, I could have been a therapist.