Como implantamos um agente de IA para operar na Polymarket

Uma sessão do Claude Code prestes a rodar um loop que caça ideias de negócio

Trecho de “O que quebrou”

Nas três primeiras semanas, o agente produziu quarenta ideias por dia, e todas as quarenta eram lixo. Ele só ficou útil quando proibi que propusesse qualquer coisa para a qual não tínhamos canal de vendas.
Antes O que ele faz Como é construído Os números O que quebrou O que eu mudaria
Custo por mês
€310
Custo de implantação
€1,200
Ideias geradas
3.100
Chegaram a protótipo
4
Receita em seis meses
€12,400
Tempo rodando
214 dias
  • Make-me-rich-SKILL.md 4 KB
  • example-idea.md 2 KB
  • sources.yaml 1 KB

Cada arquivo é verificado automaticamente contra prompt injection e conteúdo malicioso

  1. Mês 1

    Lançado com oito fontes, com as ideias lidas uma vez por semana, às sextas.

  2. Mês 3

    Cortamos o acesso ao G2: nunca gerou uma ideia boa e queimava tokens sem parar.

  3. Mês 7, agora

    Restam quatro fontes, a conta fica abaixo de €310 por mês, e nada deu errado em produção até agora.

O que as pessoas costumam perguntar

  • Quanto custou de verdade, contando o tempo de desenvolvimento?
  • Que parte você não faria de novo?
  • Isso aguentaria o nosso volume?

As masterclasses são conduzidas por profissionais de verdade que implantaram agentes nas próprias organizações.

Limpar

Mostrando 4 de 4

Vídeo 26:28

Claude apologised and promised never to do it again? Do not trust promises, set up hooks

Claude apologised and promised never to do it again, and a couple of days later it all happened again. A promise lives in one conversation, and Anthropic's guide calls the rules in CLAUDE.md advice. A hook is a check Claude cannot skip: Claude Code runs it itself, before an action, after it or at the end of an answer. Where a hook is needed (deleting data, emails, putting code online, spending money), where it only slows the agents down, and the author's five working hooks: dangerous commands, emails in his name, the task log, the skeptic and a sound. The recording of the masterclass and the working kit of files that goes with it.

Slava Sazhin 30 de setembro de 2026
Vídeo 28:22

Giving your child the right access to top AI. Building a custom web app

A child needs their own helper, not a shared chat under your login. A web app on your server goes on the phone's home screen, and Claude answers on a separate subscription. The character says what the helper is like, not what it must not do: it does not shut down a topic beyond the child's age but answers briefly and calmly, sends the child to adults about health, medicines and money, and on anything dangerous puts a mark the child cannot see, and you get a message. You read the chats, and the child knows it. The recording opens with the new Opus 5.5: which effort level for which role. The recording of the masterclass and the working kit of files that goes with it.

Slava Sazhin 23 de setembro de 2026
Vídeo 26:40

Which model for which task, so the bill does not ruin you? Do not ask Fable 5 for a casserole recipe

There are two knobs, and usually only one gets turned: the model is the ceiling of ability, and the effort level is how thoroughly it works. Slava's own blind test and the independent FrontierCode measurement say the same thing: the line of the model matters more than the effort inside it, and max costs four times what medium does without adding accuracy. Where the model and effort pair is set, and why it belongs in one file. And a fallback for helper agents when an expensive model runs out of limit: a hook catches the refusal and moves the role to a fallback model. The recording of the masterclass and the working kit of files that goes with it.

Slava Sazhin 17 de setembro de 2026
Vídeo 37:22

And who are the judges? Thanks to AI judges, Claude stopped deleting my databases

We give an agent more and more freedom, and one day it does the wrong thing. A list of bans does not save you: a ban knows words, not meaning. So in front of a dangerous action we put another Claude, a judge. It did not see how the first one reasoned: it gets only the intent, what is going to be done, why, and by which exact command line. There are several judges, they answer separately, and one vote against is enough. The permission is narrow: fifteen minutes, one thing, one use. The recording of the masterclass and the working kit of files that goes with it.

Slava Sazhin 9 de setembro de 2026

Não sabe como um agente faria uma tarefa sua? Sugira como case

Escolher um plano

(a partir de $11)

Fale com a equipe do Combobulating