#7 — Why I Built a Dev Dashboard Instead of Writing Tests

You can have the cleanest architecture in the world. If you can't feel it working, you'll never trust it.

The Mood

Six PRs in. Outbox pattern landed. Ollama integration wired. The pipeline works — I know because the logs say so.

But logs are not feeling. Logs are scrolling text in a terminal. You squint, you grep, you piece together what happened. It's forensics, not feedback.

I was tired. The event bus over-engineering burned me out (devlog #6). An AI reviewer told me the project is all architecture and no proof. He was right. We had clean abstractions, retry policies, graceful shutdown — and zero visible output. The system did things, but I couldn't see it doing things.

The Terminology

There's a concept in software development called inner loop — the cycle between making a change and seeing the result.

Edit code → Build → Run → See result → Edit code → ...

The shorter this loop, the faster you learn. The faster you learn, the better the product. Game developers figured this out decades ago — they build editor tools, debug overlays, in-game consoles, parameter sliders. Not for users. For themselves. To feel the system while building it.

The formal name is Developer Experience (DX). But DX sounds corporate. What I'm talking about is more primal — it's the difference between driving a car with a windshield and driving one with only a rearview mirror.

The Decision

I needed to see issues flowing through the AI pipeline. Not in logs. In a browser. With my prompt. Right now.

So instead of writing tests, refactoring the outbox reader, or designing the broker architecture — I built a throwaway dev dashboard.

go run . dev
→ http://localhost:8080

One HTML file. One Go file. Embedded in the binary. No npm, no React, no build step.

What It Does

Two modes:

Batch — hit "Replay & Run", watch every synced issue get classified by Ollama in real-time. Cards appear one by one. Yellow while processing, green when done. SSE streaming, no polling.

One-shot — pick any issue from a dropdown, write a custom prompt, hit "Analyze". See the AI response for that one issue. Change the prompt, hit again. Instant iteration.

The prompt persists in localStorage. Reload the page — it's still there. Change it, replay, compare. The inner loop is seconds, not minutes.

What I Learned in 10 Minutes of Playing

Things I would never have noticed from logs:

1. qwen2.5-coder:7b is too polite — it doesn't hold character. You need aggressive few-shot examples to make it play a role.
2. Long issue bodies slow everything down — some issues have 5000-character markdown bodies. The model chokes. Future optimization: summarize first, classify second.
3. The outbox has duplicates — when the poller re-syncs an issue that was already synced, it writes another outbox event. Not a bug (idempotent ack), but visible waste.
4. The prompt matters more than the model — switching from "classify this issue" to a noir detective prompt completely changed output quality. Not because the model got smarter — because the prompt gave it a voice.

None of this shows up in unit tests. All of it shows up in 10 minutes of clicking buttons.

The Anti-Pattern I'm Okay With

This code is ugly. Raw SQL in HTTP handlers. Inline HTML. No separation of concerns. The linter caught 11 issues on the first run.

I marked every file with // Throwaway PoC — NOT production code. See #152.

Because the point was never clean code. The point was shortening the inner loop until I could feel the system working. And it worked. In one afternoon I went from "logs say it works" to "I can see it working, and here are four things to fix."

The dashboard will be deleted. The insights won't.

The Principle

Build tools for yourself before building features for users.

Not always. Not for everything. But when you're in the "does this even work?" phase — when the architecture is ahead of the proof — stop architecting and start seeing.

The best code I wrote today wasn't the Ollama client or the outbox reader. It was the 200-line HTML file that let me play with both.

---

PR: #153 — the one where we stopped reading logs and started watching.