From the Backlog to the Build Log

Week of August 8 - August 14, 2025

A note about this build log entry: I published this after the week it covers. DevNitro reconstructed a draft from work-related entries in my journal, Git commit history, and time tracking. I reviewed, edited, and approved the final post. The story reflects what I knew and built at the time.

This is the first build log. This is the week I decided to revive CodeSlinger. GPT-5 landed the day before this window opened. I spent a few days testing what it could do, and by Friday I had a prototype back on the bench and a funding decision made.

CodeSlinger started out as a Visual Studio extension I started in January but left alone since March.

What I Worked On

Two threads ran side by side this week: money and code.

The money thread was a cost model. I took a console app I had been using to run the numbers and had GPT-5 convert it into a Blazor site so I could see the projections change as I moved the inputs. The question underneath it: bootstrap this out of my own pocket, or bring in a partner and raise. I ran conservative user growth against an 8x ARR valuation, tied the growth curve to the funding amount, and read the result. The partner route is far more likely to survive, so that is the current plan.

The code thread was the prototype. I pulled up the Visual Studio extension I hadn't touched since March and wired it at an OpenAI model. Then I started dragging the pieces into one place. By the end of the week the console app, the VS extension, a new CLI, and the web app (generated from my BlankSite codebase) all live in a single git repo.

What the AI Did Well

GPT-5 impressed me. Converting the cost analysis console app into a working Blazor site was fast and mostly correct. A task that used to eat a couple days of boilerplate was a short conversation instead.

The bigger tell was watching it reason about architecture for deploying AI-generated apps into real containers. That is the whole premise of the product, and the model kept up. Enough that I started believing the thing could be built.

Smaller wins stacked up too. It cleaned dead code out of the BlankSite codebase and stripped the old FlowNitro logo without me hand-holding it.

What Needed Correction

Not everything was clean.

Earlier in the week I lost almost two hours to a browsing bug on the Blazor demo components. The AI could not see it and neither could I at first. The culprit was a single extra comma in appsettings.json. No model saves you from that one yet. You read the config yourself.

The prototype is not real yet. The CLI calls out to a local Ollama model and holds a chat, but it does not print anything to the console. The proof I need is MCP working end to end so the CLI can call Daytona and spin up a container. That is next week's problem.

I also burned time I should not have. When the AI worked and I waited on it, I drifted to a phone game. This kind of build has real waiting baked in, and I need to fill it with the next task instead.

What Shipped

Nothing hit production. I doubt it will for a bit. What shipped was structural: four surfaces (VS extension, VSCode, CLI, web) in one repo, a cost model I trust, and a funding decision.

On the side I did some AI security tinkering. I ran prompt-injection experiments against Gandalf at lakera.ai and built a small image-gen API that returns a transparent pixel while logging what a custom GPT sends it. Not product work. Good for understanding how these systems get abused.

The Numbers

Category	Hours
Product / Features	41h 26m
Content	5h 49m
Community	1h 7m
Deploy / Infra	0h 25m
Total	48h 46m

More than half the week went to the cost model. I guess an early week looks like this. Before you write the app, you should decide whether it is worth writing and how to pay for it.

That is week one. A decision, a foundation, and a prototype that chats but cannot yet deploy. Back at it next week.

Curious how this site was built?
Watch the 5-minute walkthrough at devnitro.com →

August 15, 2025 Build Log: GPT-5 and the Prototype