Davide

Tech lead mostly debugging humans, systems, and now AI agents.

[Davide @iltroiani]

That means you cannot switch back to youtube while it works…

[Davide @iltroiani]

Tip: read the reasoning and steer with follow up messages while the agent is working. Avoids drifts or misinterpretation that may explode 🤯

[Davide @iltroiani]

@YashHustle_22 Yes but i had to tune a whole framework to make it sustainable…

[Davide @iltroiani]

Built on get-shit-done. Now i’m preparing a 5-min-demo that you can test on your own. Watch the repo to get notified.

[Davide @iltroiani]

What if you could just review agents’ code as if they were human?

There you go, prepare for scoped PRs and full github integration.

https://t.co/VNq8tiFRIB

[Davide @iltroiani]

Btw I think it’s useful to share my vision and let visitors understand the direction and what’s going on.

[Davide @iltroiani]

Opening issues and PR in my own open source repo.
Is it how it should work at the very beginning? 😅

[Davide @iltroiani]

@OpenVPN @FrancisDinha @ForbesTechCncl Interesting article. Orgs should give the right boundaries to avoid drift and hallucination resulting in hidden security flaws in their software.
That means working on the whole AI-assisted SDLC instead of CI/CD pipelines only.
Creating smaller reviewable diffs helps even more.

[Davide @iltroiani]

@wooing0306 Concrete results?

[Davide @iltroiani]

@riverthink Each approach has its own benefits. It really depends on the goal of the session. The default changes with the context too.

[Davide @iltroiani]

@curi0usJack Any thoughts on the implementation step?

[Davide @iltroiani]

@Taniyatweets_ You can burn either without optimized flows…but Cursor gives you room to learn more

[Davide @iltroiani]

@KataevVictor A revised version of get-shit-done that turns plans into scoped tasks on github issues and pull requests after implementation (avoiding giant diffs).
I personally worked on it for a few side projects.
I’ll leave the link if useful: https://t.co/rUBt9Z1T9L

[Davide @iltroiani]

@SamErde What framework are you using?

[Davide @iltroiani]

@SamErde If it happens on the execution phase (an agent takes over the development after the planning), then yes it’s likely. I ended up making my own version of the spec-driven framework to solve this issue. And also solved the giant diffs issue…

[Davide @iltroiani]

My preferred rule:

1 task → 1 GitHub issue → 1 branch → 1 PR → human review

[Davide @iltroiani]

A 2000-line PR generated in 12 minutes is just new debt. Watch your agents

[Davide @iltroiani]

Most of times is bad boundaries and execution scope

[Davide @iltroiani]

@GithubProjects I like his presentations at least as much as Steve Jobs’ ones

[Davide @iltroiani]

Repo link if useful https://t.co/ybVBqf7bDY

[Davide @iltroiani]

This makes spec-driven development controllable. And uses @github by default.