Benjamin Guibert

💻 Tech Builder | 🔬 Science | 🧠 Deep Thinker | 🎨 Art
🚂 Building @quotarailai

[Benjamin Guibert @bgu604]

@MarkusDerFreie I don’t use it only for dev but really with all agents I use, so it’s applied at any level

[Benjamin Guibert @bgu604]

New rule for all my agents, Claude Code and Codex included: challenge everything, my requests too.

If they think it's a bad idea, they say why in 2 lines and suggest something better. Only once though. If I still want it, they do it 🫡

[Benjamin Guibert @bgu604]

I was tired of my agents posting cron errors and me asking "what happened?" every time.

Now a script checks their jobs every 15 min, no LLM. Only a real failure wakes the agent: it fixes it, or sends me one message.

36 jobs stopped spamming me.

[Benjamin Guibert @bgu604]

Agent threads I can't handle yet were cluttering my Discord list, so my agents built the snooze from @theo 's @t3dotcodes app 😎

/snooze hides the thread. It comes back:
later → 18:00
tomorrow → noon
after reset → when my Claude quota resets

Then the agent posts where things are.

[Benjamin Guibert @bgu604]

Finished threads kept piling up in my Discord list. Now they close themselves: when an agent is done, I reply "ok", "great" or "merci" and the thread disappears from my list. No LLM call, just a word list.

Took a few fixes today because "okay" kept getting a reply 😆

[Benjamin Guibert @bgu604]

My AI agents kept fixing the symptom and making the same mistake 2 days later.

Now when one messes up, I ask "Did you learn from this?"

That phrase makes it find the cause and change a skill, a rule or a script so it can't come back.

"I'll be more careful" doesn't count 😆

[Benjamin Guibert @bgu604]

@theo up @theo 😁

[Benjamin Guibert @bgu604]

@theo could you add the possibility for agents to snooze themselves in T3 Code: it's particularly great when the agent schedule an automation for later, it could snooze itself until then. 🙏🏻

[QuotaRail.ai @quotarailai]

Claude Ultra ($200/mo) dropped the five-hour limit. Not the weekly one.

Past the weekly cap, you pay API rates on top of the $200. Ultra's $200 in API credits don't cover it.

Paying API rates anyway? Put heavy users on API keys and see spend per person.

https://t.co/PAlGyBz8XH

[QuotaRail.ai @quotarailai]

Opus 5: $5 / $25. Opus 5.5: $4 / $20.

Same 200M in + 20M out: $1,500 → $1,200 at list price. Anthropic says ~40% less on typical work.

Did your Anthropic bill move?

https://t.co/9zXwZrQMfI

[QuotaRail.ai @quotarailai]

A dev set a $30 OpenAI spend limit. A leaked key ran up $285 in an hour anyway.

The limit was saved, but "Enforce a hard limit" was off. Without it, the limit never stops a request.

Check that toggle today. Then give every key an owner.

https://t.co/BB2Al3TmsL

[QuotaRail.ai @quotarailai]

If your code says gpt-5.6, it runs on GPT-5.6 Sol. The alias routes there.

OpenAI says Sol's promo price ends Nov 30: output $20 → $30 per 1M tokens, input $4 → $5.

Find out which keys call it before December's bill does.

https://t.co/5vn5FeU2sc

[Benjamin Guibert @bgu604]

@RosoAI Viens à Chiang Mai, c’est bien 😎

[Benjamin Guibert @bgu604]

@robj3d3 Hello from Chiang Mai 😎

[QuotaRail.ai @quotarailai]

OpenAI's new 12-hour prompt cache: a write costs 2.5x input, each read ~0.1x.

Send the same prefix twice and you pay 2.6x for 2x worth of tokens. It only wins from the 3rd send.

Long system prompt hit all day? Turn it on. Anything else, measure first.

[Benjamin Guibert @bgu604]

@iziatask @VengeonsP Can I steal ChatSEO traffic? 👀

[QuotaRail.ai @quotarailai]

A Claude Max subscriber scheduled claude -p runs with ANTHROPIC_API_KEY set in their shell.

$1,800+ in API charges in two days, on a separate account. (claude-code GitHub issue #37686)

Before an agent runs on cron, know whose key it's spending.

https://t.co/kNk8GPBdxo

[QuotaRail.ai @quotarailai]

Claude Haiku 5.5 is 90% cheaper than Haiku 4.5. If your prompt stays under 100K tokens.

Above 100K it's 50% cheaper. Anthropic's own average estimate: ~75%.

Agents that stuff long context into every call should run the math on their own traffic, not the headline.

[Benjamin Guibert @bgu604]

@HermesWatcher I actually never have Hermes open, I only use Discord as an interface with it

[Benjamin Guibert @bgu604]

@mattpocockuk @itsindishere The quota is quite decent now on Claude Code, I really need a huge amount of threads in parallel in order to reach it.