Hi folks, thanks for sticking with me through three weeks of The Blueprint. This week we’ll be talking about the tools I use on top of Opencode + NVIDIA Inference Microservice, as well as some thoughts on how companies are using AI and what might be coming next.

Seems like big tech companies are finally starting to realize that pushing AI anywhere and everywhere does not actually increase productivity on its own. I’ve heard that in a lot of cases software engineers were told that they could use AI or get fired - so many decided to just run agents on dummy tasks in the background all day, burning tokens.

So why do I bring this up?

a) it’s hilarious.

and b) we, the little guys, get to learn from the big guys’ mistakes without paying millions of dollars.

I take a utilitarian approach to AI and what it can do, but also how this relates to business outcomes for SMBs. We measure cost vs value, and if it’s not working, we figure something else out.

I learned a lot about this in a previous role of mine. I worked on a safety project for a mining company creating training videos. Securing buy-in from mine workers was obviously part of the production design, but at the end of the day it wasn’t my responsibility as the person creating the product - it was the responsibility of the Change Manager. This is a role that no one envied and that everyone agreed was the hardest position on the team.

This point of this whole job was to talk to workers and explain to them why they needed to get on board with this new initiative. They hosted tool talks, group sessions, met with managers, explained how they could help out too, etc. Every step of the way was difficult and met with resistance - and this is a safety project, literally a matter of life and death!

A lot of industries know that change happens at a glacial pace. Tech, being the disruptor and general advocate for change, is learning that it’s not that simple.

I do think they’ll get through the growing pains eventually, but it presents an opportunity right now. We have access to great tools being subsidized by this push for adoption, so we might as well use them to build things that are useful!

This is the last part of that ‘car’ metaphor - where Opencode is the chassis, and NIM models are the engine. The accessories are, well, accessories - like a GPS system! Or I guess Google Maps if I’m trying not to sound old. You can do without these, but it makes it a lot easier to get where you’re going.

I built deepanswer to work on top of the other two layers. It is a multi-agent tool that discusses your query through several perspectives then synthesizes a better answer. So the whole stack works like this:

This is a lot more than just a chat window, and the addition of deepanswer to this flow is super helpful for getting verified data that is more helpful than the surface level answers you get from AI chats. Building a project in Opencode also means that you have an evolving context that the agent can learn from and add on to.

There are a lot of ‘accessories’ for Opencode out there, more often referred to as Skills or Commands. I am always testing out new ones, adding stuff, dropping stuff, etc - so I think I have a near endless amount of repos to talk about. I’ll mention about a couple others in the next section!

We’re focusing on agent skills today! These are two that I’ve actually used, so let’s get started.

This coding method is almost a year old now (wow!) but it’s still an interesting one, and it integrates seamlessly into Opencode here. The ‘ralph’ philosophy is to just have your agent stumble, fail, and bang their head against a wall until they get it right. It’s basically a loop that only exits when successful conditions are met.

This is sort of cheating because it isn’t actually a skill - this is a CLI tool that hosts that for loop. But it seems to work pretty well - again this is sort of overkill for my purposes but it’s worked when I’ve used it.

Grade: B 👍
Works but not token efficient, overkill for basic stuff

I do think this one is super widely used but it’s worth mentioning here! This one plugs in more cleanly to Opencode and uses its native skill feature. This basically levels up the planning for your agent - making sure that you are involved at all steps in the process, then chunking down that plan into its most basic forms and with best practices for development. It’s great!

My only criticism is that I think the planning process is actually a bit verbose for my workflow - as someone who likes to set their agent, go do something, then come back, it can be a bit frustrating to see that they’ve just asked another question that I basically have already answered!

Grade: A- ✅
Very good, a bit verbose but I guess that’s the point

The Blueprint Edition 3 is in the books. From here we’re just gonna be talking about AI news and repos of the day, as well as any projects I’m working on or new career developments. Thanks for tuning in and see you next week :)

Don’t get me wrong - it can be super overwhelming thinking about all the AI tools that are out there, which ones are best to use, and what will actually help you automate or optimize your workflow. I’m always open to discuss software or help you build a custom system!

Feel free to reply to this email with any questions or particular work problems that you want to automate :)