Do you use any other automation besides prompts? I try to put deterministic steps in tools as prompts are not as reliable as a script, for example.

As usual, if you need any help feel free to drop me a message and I'll do my best to support you. Have fun!

You can check the source code for AIDA v3 at github.com/danicat/aida.

If you want you to play with A2UI on your own, I highly recommend you start with a simpler use case though, and for this the codelab codelabs.developers.google.com/adk-gemini-a2ui is a good starting point.

Most engineers don't recognise code reviews as part of the job, but it is one of the most important parts.

The only thing worse than reviewing AI-generated code is reviewing AI-generated code that not even the author cared about.

Don't be that person. Future you and your team will appreciate it.

To stop the snowball without burning out:

1. Automate the pedantic reviewer: linters and tests to catch sloppy code.
2. Review what matters: core business rules.
3. Fix secondary code by sampling.
4. Continuous improvement: refine agent skills and prompts after each session.

Now hand that to an agent. Agents have zero taste and assume existing code represents acceptable practice.

Pre-trained on public code, they already skew mediocre. Stop reviewing, and old code encourages bad practices, quickly snowballing to unmanageable levels.

The broken glass was the trigger. It signaled that nobody cared, so collectively people stopped caring as well. The exact same thing happens with code.

Once there is a broken window (typos, bad names, poor tests, wrong abstractions), it triggers our psychology to stop caring, snowballing over time.

This problem always happened with human code, but much slower. Some tried to stop it and were deemed pedantic reviewers: loved by few, hated by most.

In 1969, researchers left an abandoned car untouched for a week. Then they smashed one window. Within hours, passersby stripped the vehicle bare.

Finally, I compare operational runtimes: Google Cloud Run and the Gemini Enterprise Agent Platform.

📄 Read the full article and runnable code on the blog:
danicat.dev/posts/gemini...

1. Go GenAI SDK (google.golang.org/genai): Low-level API mapping. Great for CLI tools and single binaries.

2. Genkit Go (genkit.dev): Flows and Tools, struct schema reflection, and built-in OpenTelemetry.

3. Google ADK (adk.dev): Multi-agent hierarchies, runners, and session resumption.

Read the full guide and explore the open ecosystem:

📄 Article: danicat.dev/posts/the-pr...
🌐 My personal Agent Skills catalog: skills.danicat.dev
🥋 kungfu CLI: github.com/danicat/kungfu

To keep skills sharp and avoid token bloat:

• Refinement: Use /skill-optimizer for token budgets
• Agent Dreaming: Run /learn in Antigravity to extract recurring fixes from past sessions
• JIT Streaming: Load skills on-demand with the pure Go kungfu CLI to keep global context clean

I classify skills into 3 levels:

1. Docs: Injects modern, idiomatic code when models hallucinate obsolete APIs.
2. Process: Bundles deterministic scripts/CLIs with instructions for repeatable chores.
3. Traits: Shifts cognitive posture (/grill-me, /double-diamond, /uno-reverse).

Skills are a fairly new standard, but they've become so ubiquitous that it's hard to believe that just one year ago they didn't even exist.

In my new article, I explore why the "skill of creating skills" is the most important skill a developer should master nowadays (pun intended).

Read the full chapter and runnable Go code walkthrough on the blog:

danicat.dev/posts/gemini...

#Golang

1. Bare Go GenAI SDK: 1:1 wire protocol with explicit loop control.
2. Genkit Go (genkit.dev): strongly-typed Flows with automatic schema reflection and OpenTelemetry tracing.
3. Google ADK (adk.dev): modular multi-agent orchestration with decoupled session persistence.