Fernanda Kipper

@SimInSilico im testing gemini.google/overview/agent/ and Claude Cowork I use just for normal tasks automation, like form filling, DM messages, clean up my gmail, tasks that need to combine multiples websites, simple scrapping… things like that

@kipperdev Solid stack. How are you using Gemini for browser automation specifically? I have been doing a lot with Playwright + CDP lately and curious how the AI-native approach compares for reliability.

tools I use to run my startups in 2026: 1. General Tasks + Browser Automation - Gemini - Claude Cowork 2. Sales - TheVibeCRM - LinkedIn Sales Navigator 3. Analytics + Observability - Posthog - New Relic 4. Ops - Notion - Slack (Linear + Cursor integrations) - Linear 5. Design - Figma (+ Figma MCP) - Stitch - UX Copilot 6. AI - Claude - Cursor 7. Content Creation - Super X - Gemini + Claude - Pora

@dlemesdev usepora.com.br 💜

@kipperdev what's PORA ?

@MingtaKaivo exactly! i have been deep in the trenches optimizing agent workflows and refining system prompts lately. The shift from micro-managing code to macro-managing agent logic is real

@kipperdev The real skill now: prompt engineering at scale. Been there. Went from obsessing over lint rules to writing system prompts that shape how agents think. ROI is wild.

The role of software engineer is changing completely AI will write most of our code soon, and it'll also read it for us We need to focus more on the big picture and efficiency, not syntax 1. Stop obsessing over perfect syntax and patterns 2. Start thinking about system architecture 3. Focus on business logic and the real user problems 4. Learn how to prompt AI properly (and combine multiple AIs to help you) 5. Get good at reviewing and optimizing code AI handles the repetitive parts of code We handle the strategic and creative thinking

@knutar_io 100% agree

@kipperdev This is what knowledge tools should do natively. Ingest a book, chunk by concept, surface the right piece when context matches. Right now it's manual CLI work. The tool that automates this loop - raw reference to context-aware retrieval - wins the next round of PKM.

best Claude Code tip you'll see this week: - create skills from books to make it more specialized in subjects 1. dump the PDF to MD: Use marker or pandoc to turn that PDF into Markdown. 2. Slice the MD: Use a simple script to split the MD by headers (##). One file per chapter, and store it in a ref/ folder. Claude loves small, bite-sized files. 3. the trigger: In SKILL.md, write a killer YAML description: "Use when I ask about [Topic] or [Specific Pattern]." -> @claude will now "wake up" the book only when it smells those keywords. 4. the instruction: in your SKILL.md instructions, add: "If the user asks for a specific pattern, use "grep -r" on the /ref folder to find the exact chapter, then read_file to digest it. watch the magic happen.

@MarescaAlexis try it, you won’t regret it i have been combining refactoring and UI/UX principles books with Claude Code, and the results are honestly on another level, it delivers a quality of code I’ve never seen an AI produce before

@kipperdev been doing this with technical docs but never thought to use grep to trigger the reads that's actually genius

yep! I’m using the Design Patterns book from Refactoring Guru for all the refactoring in my codebase, and the result is unbelievable!! also, we are using some UX/UI books to improve micro-animations and frontend experiences. The usability of the AI-generated components is awesome, those tiny details really level up the UI, seriously

@kipperdev can you give a real example from your daily use?

@giordanorandone totally agree

@kipperdev Knowing what to build and why. Asking the right questions. That's what will matter.

@codingandcatss if this happen, I’m scare of this future… cause what’s left for us?

@kipperdev I believe, strategic and creative thinking will be also done by AI soon.

Eu fico alternando entre Claude e Gemini Gosto do Claude pq ele tem o modo Coworking, aí ele literalmente consegue fazer umas coisinhas pra mim muito além de só responder (tipo peço ajuda as vezes pra organizar uns e-mails, arrumar uns arquivos q deixo solto, montar umas planilhas etc) Gemini eu gosto pq acho as respostas precisas e ele já tem um histórico grande meu e sabe já me ajudar a escrever uns textos mais do meu jeito

cada vez mais venho flertando com a ideia de usar o claude tambem para chat ao inves do GPT quem ja fez essa mudança, o que ta achando?

Eu fiz uma reserva de emergência pra poder tomar essa decisão, mas simm, eu entendo, eu tenho pra quem pedir socorro caso tudo de errado e entendo o ponto! Só acho que empreender não é um privilégio para poucos, acredito que tem um custo de oportunidade mais baixo para alguns, mas todos podem tentar (só que alguns estarão correndo mais risco)

@kipperdev @sseraphini Não é tão simples, você tem apoio caso alguma coisa de errado? Nem to falando de apoio financeiro mas poder voltar a morar com pais, não ter pessoas que dependam de você (não digo só de filhos mas familiares mais velhos), querendo ou não, empreender é um privilégio para poucos.

foi isso que me fez ter coragem e pedir demissão:

@dlemesdev Omg I had no idea about like and liked the comparison with pentesting techniques… I was being naive and just thinking in “allowed” ways to do Anyway, thanks for the comment! It was exactly the type of comment I was looking for <3

I used to scrape more than 100 million songs per week, in addition to albums, artist profiles, and other data from Spotify. I did this for about four years. I even suspect that the same technique used by Anna’s Archive to download the entire Spotify catalog was the same one I used. Well, high-volume scraping is like pentesting (in my view): it’s a cat-and-mouse game. In Spotify’s case, during the first year I found a link that returned a token with no query limit. That link was blocked a little over a year ago; now if you access it, a message appears saying it is for internal use. When that link that returned the token to query the APIs stopped working, I started looking for another way and noticed that the anonymous token they provided on the first visit to the site worked to query the internal API. So my flow became: open Playwright, capture the anonymous token, and then I could query all the APIs freely for 1 hour. I used around 30 threads in each of the 15 workers, which were AWS ECS Fargate Spot. At first they ran for about 6 hours, but I noticed the error rate increased even though I renewed the token every hour. So I started killing all the workers and spinning up new ones; that way there was IP rotation and everything worked very well. After Anna’s Archive, they blocked this loophole that allowed access to internal APIs using that token. Shortly after that, I left the company. I also did scraping for other streaming platforms, such as Deezer and YouTube Music. I also extracted data from ECAD. That one was annoying, but the technique was interesting: I used disposable emails from Temp Mail. So I would register, validate the account, and extract data until the account got blocked. Then I created another one and repeated the same flow. All of this was fully automated with no human interaction. It was fun 🙂

Anyone knows how do the API scrapers work under the hood? I mean, i know how web scraping works - requests, parsing HTML, seletores CSS... But how do they manage to scrape the social media profiles without having their IPs blocked? Rotating proxies? Residential IPs? Headless browsers com fingerprint spoofing? And get always updated content. Rate limiting must be a nightmare.. Do they use webhooks? Smart polling? Some kind of distributed cache? Curious about the infra behind services like Apify, SearchApi, Bright Data etc etc etc

@pedrobarretto_ @BolhaDevs @VictorTaelin A partir de 50% já começa a piorar aqui Tem uns cara que fizeram um script zinho pra minorar a janela de contexto do Claude, vale a pena olhar reddit.com/r/ClaudeAI/com…

Até quantos % da janela de contexto preenchida uma LLM consegue gerar um resultado bom? Depende do modelo? Cc @VictorTaelin

@maincarmem @thesanchesdev @daniellimae @ChristoPy_ @sseraphini Tem essa ferramentinha aqui tbm que é legal de usar pra testar se vai cair em spam glockapps.com/inbox-email-te…

@thesanchesdev @daniellimae @kipperdev @ChristoPy_ @sseraphini vai salvar muito!!!! obrigadaaaa

recebi algumas mensagens dizendo q o nosso e-mail de acesso a caramelo a acabou indo pro spam… sendo assim, tomara que o algoritmo alcance todo mundo que acabou passando pelo mesmo! cc @daniellimae @kipperdev @ChristoPy_ @sseraphini

@maincarmem @daniellimae @ChristoPy_ @sseraphini caiu no spam pra mim tbb :/

Eu concordo realmente a gente ta numa era da busca pela produtividade extrema e até alguns umas semanas atrás eu não acreditava em conteúdo gerado por IA Porém depois de ver alguns cases de empresas que estão usando isso pra trazer tráfego pro seu site e para trazer leads eu não podia deixar de tentar…. Quando a gente tá no começo da empresa e tem pouquíssimos funcionários qualquer ferramenta de produtividade que ajude a delegar tarefas e permita que a gente consiga fazer todas as coisas são muito essenciais Então esse agentezinho que eu estou criando não é pra delegar a minha criação de conteúdo mas sim pra tentar criar um novo canal de distribuição pra empresa. Eu vou começar com conteúdos textuais e depois eu quero ir avançando pra conteúdos de vídeo… vamo vê se funciona

Sei lá, esse negócio de conteúdo gerado por IA nao me pega. Se fosse pra ser assim, o Instagram poderia ser um ChatGPT que mostra vídeo gerado na hora. Eu compreendo que na era da IA existe uma busca explícita pela produtividade e pela eficiência. Mas eu não consigo notar que realmente fizemos progressos bons e saudáveis, apenas que estamos em um lugar diferente agora. No fim queremos que a IA faça exatamente o que tememos que façam conosco: trabalhe 24/7.