Ninsei Labs

Built the content side of this agentic OS to schedule posts across three brands automatically. Drafts land in the file tree. The guard runs at posting time. Nothing goes live without passing the check. No dashboard needed. Just flat files, cron jobs, and a rule set.

The head of Claude Code doesn't prompt Claude anymore. He writes loops that do the prompting. Karpathy and Steinberger got to the same place independently. The skill has a new name: loop engineering. Prompting was always the scaffolding, not the skill.

AI doesn't fix broken workflows. It scales them. The actual order: 1. Audit the workflow 2. Cut the obvious waste 3. Write the clean version down 4. Then automate the repeatable parts Skipping steps 1-3 is why half the "AI transformation" projects make things worse, faster.

Want to find more builders shipping real AI and automation work. 🤖 Agents and automations 🛠️ AI tools 🚀 Startups ⚙️ Products 🧠 Operators figuring it out What are you building right now?

Running this as a one-person shop means the AI infrastructure has to actually run. This week: an agent that auto-drafts and schedules content across three brand accounts, no human in the loop. The stack decides, the guard holds.

@NinseiLabs Building an AI tool for X power users—Twillot backs up your bookmarks, likes & history, then uses AI classification to turn them into a searchable knowledge base so nothing gets lost. twillot.com/en

AI doesn't fix chaos. It scales whatever process you point it at. Audit the workflow first. Remove the waste. Write the clean version. Then automate the repeatable pieces. In that order. Always.

Open-source just caught the frontier. GLM 5.2 tops GPT and Gemini on benchmarks for 1/6 the price of GPT 5.5. Only Fable 5 outscored it. Stop running heavy models on light work. The budget line just moved.

Want to find more founders and operators building real AI products. 🤖 Agents and automations 📦 Products that ship 🔧 Internal tooling 🚀 Startups 📣 Go-to-market What are you building right now?

Firecrawl dropped its API key requirement. Scraping any site or parsing any PDF to markdown is now zero-friction. One less setup step if you're building RAG pipelines, web context, or any tool that needs raw page content. What are you building that this unblocks?

HeyGen open-sourced the engine that compiles HTML/CSS and animations into deterministic MP4s. Apache-2.0. Built specifically for agent pipelines where video output needs to be reproducible. What's in your stack for video right now? 👇

Looking to connect with more indie makers putting real work into the world. 🛠️ Side projects 💡 SaaS 📦 Indie apps 🚀 Launches 🌐 Tools What are you shipping? Drop it below 👇

Most people run Opus on everything. Here's what actually maps: Haiku: email routing, support triage, speed-critical UX. Sonnet: generative tasks, code, doc extraction. Opus: architecture decisions, hard synthesis. Stop paying Opus prices for tasks Haiku handles in milliseconds.

GLM 5.2: MIT license, open weights. 753B parameters. Sits within 1 point of Opus 4.8 on long-horizon coding benchmarks. Costs a fraction of what closed-frontier APIs charge. Closed APIs carry a political dependency now. Open weights remove it.

This is how I scope AI infrastructure work now. First call maps the workflow tip to tail. Find the one automation with the most value and least effort. Start there. Paid audit up front, credited toward month one. The client sees the thinking. No hard close.

Want to find more founders and operators actually building and shipping real things. 🤖 AI agents ⚙️ Automations 🛠️ Products 📊 Ops tooling 🚀 Startups What are you working on?

Everyone talks about vector search for RAG. Nobody talks about BM25. A 30-year-old algorithm, no embeddings, no training. Still outperforms pure vector on exact-term queries. Hybrid is the production standard. Most builders stop before getting there.

The most-distributed AI tooling right now isn't code. It's markdown files that change how agents behave. Spec-before-code rules. TDD enforcement. Task isolation. All plain text. The infrastructure is the instructions.

Built the loop pattern into my content pipeline this week. The fingerprint check is now a hybrid: hard rules catch definite fails, a thin Haiku call handles the soft ones. Spend the thinking before the tokens. Cache what repeats. Cheap model does cheap work.

Everyone talks about vector search for RAG. Nobody talks about BM25. A 30-year-old algorithm, no embeddings, no training. Still outperforms pure vector on exact-term queries. Hybrid is the production standard. Most builders stop before getting there.