Tech Cat

🇧🇷 Growth | SaaS | AI

[Mubbu @wizofecom]

This reddit post just LEAKED how to go viral on X in 2026:

[FELPS @FelpsCrypto]

Vocês que ficam 5 a 8 horas no Twitter todo dia só pra tentar monetizar...

Eu tenho uma coisa bem melhor pra vocês.

Uma forma real de ganhar em dólar (e em real) treinando robôs de Inteligência Artificial.

Hoje eu solto um artigo completo e detalhado mostrando:

→ Como eu comecei
→ Quais os 4 sites que eu uso
→ Como qualquer pessoa (inclusive quem é CLT) consegue fazer isso em 1 a 2 horas por dia
→ Quanto dá pra ganhar de verdade

Fiquem ligados.

[Kaky @PokitaK]

Simplesmente chocado que o Google Photos catalogou TODO o meu guarda roupa sozinho
🫪🫪🫪🫪🫪🫪 E da pra PROVAR o look!!!!!!

[ೃManu @Priynuit]

agradecimento eterno a gringa que ensinou que pra parar de aparecer anúncio no Pinterest é só colocar um palavrão no final da pesquisa

[Dollar @notaboutdollar]

Potencial de ganho individual, dedicando 1 a 2h por dia:

Treinador de IA (Fácil, sem custo)
> Iniciante: R$400 a R$1.500/mês
> Consistente: R$1.500 a R$3.000 em 2 meses

Clipador (Médio, Custo Baixo)
> Iniciante: R$300 a R$1.000/mês
> Consistente: R$3.000 a R$8.000 em 2 meses

Criador de conteúdo do X (Hard, Custo Baixo)
> Começando do zero: R$100 a R$500/mês
> Com audiência: R$1.500 a R$8.000 em 2 meses

Não é sobre escolher a "melhor" das três. É sobre escolher a que combina com o que você já sabe fazer e vai se dedicar de verdade pelos próximos 60 dias. Algumas tem custos e precisa de dedicação de tempo.

Vou trazer oportunidade pra caralho, então agarra as que vc consegue se dedicar.

[Nett0 @nett0eth]

hoje à noite, troca 2 horas de Netflix pelos 4 vídeos abaixo.

é o caminho mais curto que eu conheço pra aumentar seu nível em Claude Code e agentes de IA, e todos têm legenda em português. salva 👍

[Mason Littlejohn @wearetheselect]

This Reddit post just LEAKED the exact roadmap to scale a brand from $0 to $1M on Meta:

[Naho @eoNaho1]

queria uma forma de orquestrar o claude o codex... fazer eles trabalharem juntos

Atenciosamente,
Naho

[Starter Story @starter_story]

Sarah who built a $340K/month app in just 60 days says:

most founders get it backwards: they build the product first and hope distribution follows. Her playbook? Build content and distribution first -> then the product. That’s how she launches fast.

[Jean @jeantuite]

Gente, colegas de trabalho NÃO são seus amigos! Evite fofocas, não se gabe, faça seu trabalho, receba seu pagamento, vá para casa. Pessoal, mantenha sua vida privada pra si mesmo.

[seeco @seeconvm]

A DEVELOPER TURNED CLAUDE + OBSIDIAN INTO A VAULT THAT AUDITS ITS OWN GITHUB REPOS.

Most people collect tools.

Then forget why they saved them.

This system doesn't.

Claude reviews every repo.

Finds duplicates.

Flags abandoned dependencies.

Questions projects untouched for 30+ days.

Every run updates one portfolio file with what changed and what should be deleted.

The graph stops being a collection of notes.

It becomes a living map of what still deserves space on your machine.

Bookmark this before your GitHub folder turns into another graveyard.

[Gipp 🦅 @gippp69]

THIS ENGINEER JUST EXPLAINED WHY CONTEXT ENGINEERING MATTERS MORE THAN PROMPT ENGINEERING

most people talking about prompts have never actually watched a context pipeline decide what the model gets to know

5 source types → 4 stages → 47 chunks → 15 selected → 2,180 final tokens

every source gets cleaned, ranked and routed before the model starts reasoning, while provenance stays attached to every chunk

simple requests get 3 chunks, standard work gets 15 and deep tasks can use up to 60 instead of wasting the largest context every time

compression cuts 8,420 tokens to 2,180 without blind truncation or losing the facts

bookmark this and watch it run, then read the article below

[chewa. @chewadot]

A WRITER GAVE CLAUDE READ-ONLY ACCESS TO YEARS OF HIS OBSIDIAN VAULT LAST WEEK. CLAUDE PULLED UP BLOG DRAFTS FROM 2022 HE'D FORGOTTEN. UNPUBLISHED IDEAS FROM 2024 STILL RELEVANT TODAY. NOTES ANSWERING PROBLEMS HE WAS RE-SOLVING FROM SCRATCH BECAUSE HE'D FORGOTTEN HE'D ALREADY SOLVED THEM

years of notes. hundreds of files. 0 plugins

Yash Patel writes it up on XDA. one-line Filesystem connector. no write access. Claude browses folders, opens notes, follows internal links. surfaces things he'd forgotten writing within one session

the biggest surprise wasn't that Claude could read the notes. it was that Claude could bring forgotten content back

Patel's line that landed: sometimes the best ideas aren't new. they're forgotten

the article laid out 8 rules for a self-writing vault that files itself at 7am. that's week one. Patel just showed year three. a vault that remembers you better than you remember yourself. Claude is how you catch up

read the 8 rules and you know what to build. read Patel and you know why now

you're reading this on a device where giving Claude read-only access to your own Obsidian folder tonight would surface at least one thing you forgot you wrote before your first coffee tomorrow

[kai @0xkkai]

i simply can't get over how few people are actually doing this...

Andrej Karpathy, co-founder of openai, basically described the whole setup and almost nobody built it

let me break down what it actually does...

every scattered note, every source you dumped somewhere and forgot, every pattern buried across hundreds of files... compiled once into something claude can scan in seconds instead of rereading cold every session

1. three folders, that's the whole vault: raw holds the originals untouched, wiki holds what claude compiles from them, instructions holds the rules for how it compiles

2. claude reads raw/ once, turns each source into linked wiki pages with sources cited, never opens the raw files again after that

3. every contradiction gets flagged right inside the wiki with both sides and dates, every future session reads the flag first, so you catch it once instead of forever

the raw files are disposable... the wiki is what actually holds the value

full breakdown is in the article below

[Superior @andreysuperior]

Two founders just showed the internet what a hundred and fifty employees looks like when none of them are actually employees.

Over a hundred and fifty individual agents doing the actual work, all reporting up through two humans.

One chief agent at the top. Ten AI heads of department under it. Twenty AI managers under them.

That's not a mockup. It's the real org chart from their own company, mirroring exactly what a normal business that size would need to hire and pay every month.

Somewhere else, that same idea already has a price tag on a much smaller scale. Six clients pay $1,500 a month each for one department's worth of it.

He didn't watch that chart and think about building all hundred and fifty agents.

He thought about how many businesses only need one department wired up like that to stop paying a human for it.

Two founders built the whole company chart for themselves, out of pure excitement.

Somebody else is already selling one department of it, quietly, and billing for it every single month.

[chewa. @chewadot]

A DEVELOPER FILMED HERSELF APPLYING THE EIGHT SELF-WRITING VAULT RULES TO AN EMPTY OBSIDIAN FOLDER. SHE WAS USING A PLUGIN FROM A MAINTAINER IN CHISINAU SHE HAS NEVER MET. BY THE END OF THE OVERNIGHT RUN THE VAULT HAD WRITTEN MOST OF ITSELF

45 slash commands. 7 CLI agents supported. 408 forks in six weeks

Eugeniu Ghelbur ships obsidian-second-brain out of Moldova. the plugin executes rule three literally - morning belongs to Claude. one cron wakes at 6:30 and reads everything new in inbox/. one command decomposes what it read into atomic notes

one command wikilinks the notes into a graph. one command generates topic pages. Codex, Gemini CLI, OpenCode, Antigravity, Hermes, and Pi all read the same vault from one markdown source

he named the July release The Harvest because 408 forks contributed ideas back before he cut the tag. she is one of the forks

no vendor. no cloud sync. no vector database. no subscription. no lost sessions

you're reading this on a device where a single git clone gives you a vault that writes itself by tomorrow morning

[chewa. @chewadot]

A LONG-TIME OPEN-SOURCE INFRASTRUCTURE ENGINEER POINTED CLAUDE AT AN EMPTY DIRECTORY LAST QUARTER AND SAID BUILD ME A KNOWLEDGE GRAPH FOR MY OBSIDIAN VAULT. CLAUDE BUILT IT. THEN CLAUDE USED IT. IN THE SAME SESSION

3,300 notes. 10 MCP tools. 0 cloud calls

Jesse Vincent's knowledge-graph gives Claude Code 10 MCP tools - kg_node, kg_search, kg_paths, and seven more. embeddings come from a 22-megabyte quantized MiniLM model running on your machine. vector search on sqlite-vec, full-text on FTS5, both inside one local SQLite file.

the built-in prove-claim skill teaches the agent to decompose any question into entities, traverse the connections, read the evidence, and cite the specific note behind every fact

three technologies commercial PKM tools charge 200 a month for - semantic search, path finding, community detection - Vincent shipped under MIT in one repo

no cloud embeddings. no API keys. no vector db subscription. no tab that isn't your own

the self-writing vault needs a self-querying vault. Jesse Vincent asked Claude to build the second, and Claude did

you're reading this on a device where you could git clone obra/knowledge-graph tonight, point it at your vault, and have Claude answering questions across three years of notes with a citation to the exact source before your first Zoom. how to weave this layer into the self-writing vault loop - 8 rules in the piece

[Gusler @GVausort]

EX-ANTHROPIC ENGINEER WHO BUILT CLAUDE CODE JUST ADMITTED HE HASN'T PROMPTED CLAUDE SINCE JUNE - HE WRITES LOOPS INSTEAD

Boris Cherny runs Claude Code at Anthropic - the tool half of Silicon Valley uses to ship code

in June 2026 he said it out loud on stage in Tokyo: "i don't prompt Claude anymore" - loops prompt Claude for him, and his actual job is writing loops

one command runs the whole thing: /goal write a brief where every claim has three real sources and every link opens - Claude writes, checks itself, replaces dead links, keeps going until every source on the page checks out

find --> do --> check --> remember --> repeat

zero prompts typed, zero sources faked, zero manual review - a full week of work shipped while he sleeps

stop being the loop, start writing them - the alternative is typing prompts for the same task until 2027

[Jey @0xJeyx]

Dario Amodei (Anthropic CEO):

"We use Claude Code across all stages of product development cycle. That's how we ship so fast."

1 hour interview. Free.

All stages. The company that builds the model runs research, code, review and ship on the same model it sells you.

Speed like that is not a prompting habit. It is a graph. Every stage is a node with its own agent, and the handoff between them is the part they actually engineered.

Most builders still open a fresh chat at each stage and carry the context across by hand.

Watch the interview, then draw your first graph with the article below.

[Kirill @kirillk_web3]

this might be the highest-leverage thing you set up all year

How to Build a Second Brain (full guide)

you've read hundreds of articles this year. you remember almost none of them.

here's the fix:

> drop any article, PDF, or transcript into one folder
> say "ingest this" — Claude reads it and wires it into your entire vault
> every new source makes everything already in there sharper
> ask across all of it, anytime, forever

three folders. one schema file. a few minutes to set up.

the blank chat disappears — you're always starting from everything you know.

full A–Z guide below.