Guides / Working notes

The long version
of the short video.

Each guide is the companion to something I posted — the numbers, the commands, and the sources, with room to show the work. Every figure is checked against a primary source and linked back to it.

7 published

Read it,
then go run it.

Short enough to finish, specific enough to act on the same day. No introductions, no throat-clearing.

CAGE10 min read

Contain AI agents before they reach production

A defensive architecture for running capable AI agents with ephemeral compute, scoped files, brokered network access, short-lived identity, circuit breakers, and human shutdown.

ai-security
agents
cybersecurity
architecture

MOTION6 min read

Motion graphics in code: how to build reels that render themselves

A practical introduction to making social motion graphics with React and Remotion — why programmatic video beats a timeline for repeatable content, how the pieces fit together, and the design rules that decide whether the result reads on a phone.

remotion
react
motion-graphics
video-automation

BUILD11 min read

Build a WhatsApp AI assistant that can safely take action

A practical TypeScript architecture for receiving WhatsApp messages, continuing an OpenAI conversation, calling approved business tools, replying through the Cloud API, and handing difficult cases to a person.

whatsapp
openai
automation
ai-engineering

GUIDE9 min read

Claude Opus 5 vs GPT-5.6: build a router, not a model religion

A practical model-routing playbook for Claude Opus 5 and the GPT-5.6 family, including current prices, a decision table, a TypeScript router, evaluation criteria, and production guardrails.

model-routing
claude
openai
llmops

OPUS57 min read

Opus 5 vs Fable 5: the model picker is not your cost dial

Claude Fable 5 costs exactly twice what Claude Opus 5 costs, and that is the least useful number on the page. Here are both price tables, the one setting that actually moves your bill, and the prompt instruction you should delete today.

claude
llm-cost
prompting
agents

KIMI8 min read

Kimi K3: unmetered is not free — the actual cost breakdown

Kimi K3 is 2.8 trillion parameters and the weights are public. That does not make it free, it makes it unmetered — and the two words have very different invoices. Here is what self-hosting it actually needs, and the split that beats picking a side.

open-weights
self-hosting
llm-cost
inference

SITE8 min read

The Claude web-design setup: skills, taste, and a browser it can actually see

The four-part setup that gets Claude building sites that don't look generated: design skills for the rules, a taste skill for the references, Figma MCP for the source of truth, and Playwright so it can look at what it built. With the exact commands and file paths.

claude-code
skills
mcp
web-design

AI Lab & shipped work

The long versionof the short video.

Read it,then go run it.