Building an AI application is mostly plumbing: keys kept out of the repo, a plan for 429 responses, and a chunking engine for anything longer than the token limit. Notes on what keeps an AI feature running once it leaves your laptop.
Nobody trains a vision language model from a blank slate. You freeze a ViT, train a Q-Former to bridge image features into the text embedding space, and wrap the language model in LoRA. This is how those three pieces fit and where the GPU hours go.
Chaining agents multiplies errors instead of dividing work: DeepMind measured up to 17.2x amplification. Here is the math behind that, the three topologies that hold up in production, and the checklist I run before shipping an agentic system.
Vibe-checking search results is how teams end up rebuilding a backend for a system that scores worse than the SQL it replaced. What a real AI search evaluation needs: a golden set, repeated trials, and an ICC check before you ship.
AI infrastructure spending hit nearly $400 billion in 2025 against barely $100 billion of revenue. The skills that survive that correction are the classical ones: causal inference, experiment design, Bayesian reasoning, domain modeling and SPC.
A tool is one capability, a skill is the orchestration around it. Notes on writing agent skills in Python instead of Markdown, and where MCP fits when you build agents outside Claude.
Vibe coding a feature is easy; keeping it out of the technical debt pile is not. A CLAUDE.md handbook, plan mode before any edit, and a review pass that assumes something is broken.
Running OpenClaw on bare metal, skipping the system prompt, and handing it admin API keys are the three setups that go wrong most. What to do instead, with a docker-compose file and a scoped IAM policy.
Claude Skills keep instructions in files the agent reads only when needed, which cuts the token overhead of always-on MCP tools. Subagents push it further by throwing away their own context once the job is done.
Demo agents fall over in production because errors compound across steps and prompts break when the business process changes. Structured output, LLM-as-a-judge evals and OKR alignment are what get a prototype across.