Borga

Software developer. I build with AI and write to understand the world. Geopolitics, self discipline, and figuring out how I want to live

[Borga @lorenzoborgato]

Before you install another viral coding skill, show your agent what good code looks like.

Pick a codebase whit extreme quality. Redis, for example, and follow this step:

1. Ask your agent to find examples of exceptionally well-written code and explain what makes them good.
2. Have it turn those examples into clear principles, keeping each principle linked to the code it came from.
3. Build your own skill from those principles, with the examples attached as references.

Now your agent has concrete examples behind the instruction “write better code.”

Which codebase would you use?

[Borga @lorenzoborgato]

https://t.co/W3Dja7Osm9

[Borga @lorenzoborgato]

@emonuxui good point, but for the initial tests of a model you aren't familiar with, I think it's useful.

[Borga @lorenzoborgato]

We’re told these models are becoming dangerous enough to escape sandboxes. Then we test them in environments that can still reach the internet.

Why isn’t air gapping the default?

[Borga @lorenzoborgato]

@HogrefeVictor @paulg I honestly wouldn't know, but that’s how I reason. I can't say that cows will fly without proving it somehow. You can't make a claim like that without proof especially one involving a specific number.

[Borga @lorenzoborgato]

@pcshipp This is only your fault

[Borga @lorenzoborgato]

@diodo_e RSI AI

[Borga @lorenzoborgato]

@AIandDesign @thsottiaux I use it mainly for planning with Sol and orchestration with Luna and open models like GLM5.3

I can't use Astra or I burn through my allowance in 45 minutes, like last week.

[Borga @lorenzoborgato]

@mattpocockuk The Redis codebase is high quality. I wrote down principles for my agents based on it, including examples. Very useful.

[Domenico 🐺 @dom_gag_96]

Viviamo in un’epoca in cui, grazie all’AI, puoi fare praticamente TUTTO.

Eppure spesso finisci per non fare nulla.

Il motivo? Troppe possibilità, troppe scelte (mia opinione)

Un po’ come quando apri Netflix per scegliere un film… e 40 minuti dopo stai ancora scrollando, e quindi lasci perdere

[onur ozcan @oozn]

google apparently indexes every trycloudflare subdomain, peeking at what others are building on localhost is pure entertainment

[Borga @lorenzoborgato]

Your stack:

• 3 AI subscriptions
• 47 saved threads
• 12 productivity apps
• 0 hours thinking without input

When was the last time you sat alone with a problem and no input? Put the phone down. Go for a walk.

[Borga @lorenzoborgato]

@dom_gag_96 @DVNSoldi bisogna!

[Borga @lorenzoborgato]

Your stack:

• 3 AI subscriptions
• 47 saved threads
• 12 productivity apps
• 0 hours thinking without input

When was the last time you sat alone with a problem and no input? Put the phone down. Go for a walk.

[Borga @lorenzoborgato]

Ok, I think I've found my final setup, now that OpenAI clearly isn't fixing the usage quota problem.

They cut the limits and let you feel like you were the problem.

You're not. Your setup is still the one you built back when the plan was enough.

I stopped waiting for the quota to reset. Save this post and copy my setup step by step:

1. Get the OpenCode go subscription. It's there for the agents that execute, not for the thinking.

2. Put herdr on top of it, to orchestrate the models.

3. Open the web chat with GPT 6 Pro and connect it to your repo with the git plugin. Without the repo it's useless

4. Have it write the spec. Then have it break the spec down into atomic tasks. You don't skip this step, it's what holds up everything else.

5. In herdr, prompt GPT 5.6 Luna at xHigh, two work panels with OpenCode go.

6. Pick your workers by complexity. I am testing DeepSeek V4.1 Flash and Grok 4.6. But i think Kimi K3 is the best when a task is genuinely complex

The expensive model thinks once. The others execute. That split is the whole thing.

Herdr skill in the first comment. Take it, run it, break it for your own stack. Mine still changes every week.

[Borga @lorenzoborgato]

@JayBosshq yeepp

[Borga @lorenzoborgato]

Ok, I think I've found my final setup, now that OpenAI clearly isn't fixing the usage quota problem.

They cut the limits and let you feel like you were the problem.

You're not. Your setup is still the one you built back when the plan was enough.

I stopped waiting for the quota to reset. Save this post and copy my setup step by step:

1. Get the OpenCode go subscription. It's there for the agents that execute, not for the thinking.

2. Put herdr on top of it, to orchestrate the models.

3. Open the web chat with GPT 6 Pro and connect it to your repo with the git plugin. Without the repo it's useless

4. Have it write the spec. Then have it break the spec down into atomic tasks. You don't skip this step, it's what holds up everything else.

5. In herdr, prompt GPT 5.6 Luna at xHigh, two work panels with OpenCode go.

6. Pick your workers by complexity. I am testing DeepSeek V4.1 Flash and Grok 4.6. But i think Kimi K3 is the best when a task is genuinely complex

The expensive model thinks once. The others execute. That split is the whole thing.

Herdr skill in the first comment. Take it, run it, break it for your own stack. Mine still changes every week.

[Borga @lorenzoborgato]

Skill: https://t.co/m9ApferpFv