☕ on the menu · AI Explained

How modern AI actually works.

Not “what is ChatGPT” and not a list of prompts. The infrastructure: context windows, MCP, RAG, embeddings, tool calling, explained clearly enough to build on, and simply enough to learn from. These are the SMTP and POP3 of the AI era: foundational concepts you’ll still be looking up in 2030.

Explainers

Concept
What is a context window?
The model's working memory, tokens, what fills it, why it runs out, and how to manage it.
read →
Concept
What are embeddings?
How content becomes a vector that captures meaning, the engine behind semantic search, RAG, and recommendations.
read →
Infra
What is a vector database?
Where embeddings live, fast nearest-neighbour search over millions of vectors, and the backbone of RAG.
read →
Pattern
What is RAG (Retrieval-Augmented Generation)?
Grounding a model in your own documents at query time, embeddings, retrieval, and why it beats fine-tuning for facts.
read →
Concept
What is agent memory?
Why an assistant 'remembers' or 'forgets', context vs short-term vs long-term memory, and how persistence is faked.
read →
Protocol
What is MCP (Model Context Protocol)?
The open standard for connecting AI apps to tools and data, the 'USB-C for AI tools.'
read →
Protocol
What are MCP servers?
The servers you actually plug in: files, databases, GitHub, the web, browsers. The categories that matter, real examples, and how to connect one.
read →
Concept
What is tool calling (function calling)?
How a model decides to call a function, returns structured arguments, and your code runs it, the basis of agents.
read →
Agents
How to build an AI agent
What actually turns a chatbot into an agent: a model plus tools plus a loop plus memory plus a stop. The five parts, the agent loop, the security you cannot skip, and a builder that writes your agent's spec.
read →
Concept
What are structured outputs?
Making a model return JSON that matches a schema every time, vs JSON mode, and why it powers extraction and tool calls.
read →
Concept
What is an AI hallucination?
Why models state false things with total confidence, where the risk is highest, why you cannot trust the tone, and the concrete ways to reduce it.
read →
Concept
What are reasoning models?
Models that think step by step before answering, trading speed and cost for accuracy on hard problems. Chain of thought, what it is great at, where it is wasted, and the tradeoffs.
read →
Practice
What are AI evals?
How to measure whether AI output is good enough: the dataset, the scorer, the metric, and the four ways to score (exact match, rubric, LLM-as-judge, human). The verify gate at scale.
read →
Security
Prompt injection and AI security
Why a model cannot tell your instructions from text it just read, the lethal trifecta that makes agents dangerous, and the architecture-level defenses that actually contain it.
read →
Practice
How to prompt Claude Fable 5
Fable 5 is built for long, hard, autonomous work, and that shifts how you prompt it: effort as the main dial, saying less, grounding long runs, setting boundaries, and the migration gotcha that quietly falls back to Opus 4.8.
read →
Guide
How to write code with AI
A practical, no-hype method: set expectations, work around the training cutoff, manage context, plan then dictate, test everything, and know when to take over.
read →
Automation
What is workflow automation?
The trigger-to-action model behind every automation, no-code tools like Zapier and n8n, and where AI fits.
read →
Automation
AI agents vs workflows vs RPA
Three ways to automate work and how they differ, who decides the steps: you, a UI script, or the model.
read →
Automation
Webhooks vs polling
The two ways software learns that something happened, push vs ask, and the latency, cost, and reliability tradeoffs.
read →
Automation
What is AI orchestration?
The layer that coordinates models, tools, and data steps into one reliable pipeline, vs a single prompt, workflow, or agent.
read →
Automation
What are AI loops?
Giving a model a goal, a way to verify its own work, and a rule for when to stop, so it plans, does, checks, and iterates instead of answering one prompt at a time. The verify gate, when a loop is over-engineering, and a builder for your own.
read →

Tools

Library
AI Patterns Library

Design patterns for AI systems, Reflection, Router, Judge, Verification, Human-in-the-Loop. How each works, when to use it, and an example prompt.

Browse patterns →
Live · from a sister tool

When the AI you build on goes down

Everything on this page ultimately runs on a handful of providers, and when one of them has an outage it cascades to every tool built on top of it. Here is the live status of the major AI services, with the live AI dependency map behind it, from our sister tool AIDownCheck.

What’s coming

This pillar is built infrastructure-first, and the core cluster is live, context windows, embeddings, vector databases, RAG, agent memory, MCP, tool calling, and structured outputs, all cross-linked so each concept leads to the next. Next on the bench: more patterns and more client-side tools. Everything runs in your browser; nothing is uploaded. Read the blog for the thinking, or start with a context window.

A note on numbers: AI capabilities and limits change fast. Anywhere we cite a model-specific figure, it’s dated and kept in one place you can verify, we don’t assert specs from memory.