Curations

Keeping Claude Code minimal, forcing full reasoning with three settings.json lines, and investing in system architecture — a practical take on what actually moves the needle.

32 slash commands that work out of the box with Claude, ChatGPT, and Gemini – no custom definitions needed. Categorized by use case with practical combination strategies.

A structural analysis of the official sample code from a major Korean brokerage API, optimized for LLM agents and Python environments.

A look at macshot, a native open-source macOS screenshot and screen recording tool that delivers powerful features without the subscription burden.

An open-source platform that connects documents scattered across Notion, GitHub, and S3, then queries them with a local LLM. Runs entirely on-premise without external APIs.

Even after adopting agents, if the work structure itself does not change, productivity does not scale. Reid Hoffman’s conductor strategy starts by breaking that illusion.

Parsing raw legal XML means broken table structures and half a day lost on preprocessing. Beopmang is an API that delivers pre-cleaned JSON, solving that problem upfront.

A Spring Boot library for streaming 1M-row Excel downloads without OOM. A single annotation eliminates the boilerplate.

A self-hosted RAG platform that runs on Ollama to unify team documents scattered across Notion, GitHub, and S3. Natural language search works even in air-gapped environments without external APIs.

Repeating ‘fix this’ only circles within problems you already know. Deploying expert personas reveals blind spots, and learning when to call which persona also saves tokens.

When NotebookLM stops at 20 slides, most people just give up. Designate a single master script as the source and push through with code-style commands to generate 40 slides.

How assigning Claude, Perplexity, and NotebookLM to distinct stages of lecture planning changes the workflow. The point is not using more tools but knowing which tool fits which stage.

Tried extracting web data for a RAG pipeline with defuddle and found results vary wildly by site structure. On sites where semantic HTML has collapsed, body text and ads mix together, and in dynamic rendering environments the content itself vanishes.

Letting an agent handle planning end-to-end causes it to loop and stall in production. The OpenAI guide addresses this with explicit workflow control.

When you first deploy AI agents, outputs land in random folders and formats vary every time. Aligning folder structure to the org chart and codifying SOPs is what makes quality consistent.

If you have ever stalled in front of the plugin list when first setting up Claude Code, Plugin Advisor narrows that entrance. Preset Packs and pre-flight checklists catch runtime errors before they happen.

Connecting NotebookLM, Claude, and Gemini as a pipeline to turn reports into slides. A single model struggles to handle structuring and formatting simultaneously – splitting roles produces better results at each stage.

With LangChain’s move to v1, legacy chain code has become outdated. A repository of Korean Jupyter notebooks walks through LangGraph’s State and Node design step by step.

Examining an automated reporting system built with Python and GitHub Actions to escape the tedium of manually checking macroeconomic indicators every day.

Analyzing a workflow where an AI agent trained on design principles generates polished presentation materials as a single HTML file.

Addressing data bottlenecks encountered during API integrations and applying n8n’s core mechanics to real-world practice.

Examining key features of a Python-based document conversion tool released by a global IT company, and proposing efficient approaches to handling data in LLM pipelines.

korean-law-mcp streamlines legal data access while DataNexus’s knowledge-graph-based ontology layer explicitly maps complex statutory connections, significantly enhancing AI’s ability to interpret law accurately.