AI basics — a plain-language primer

1. What is an LLM, and why does it “forget”?
2. What is a context window?
3. What is Claude Code?
4. What is an agent?
5. What is memory (in aiflow)?
6. What is a skill / slash command?
7. What is a hook?
8. What is MCP?
9. What is the Ralph loop?
10. Token? Cost?
11. Where to go next

---

Most people struggle to set up their AI project successfully — usually not because the tools are bad, but because the concepts are new. This page explains them in plain language. Skip it if you already work with Claude Code daily.

What is an LLM, and why does it “forget”?

A large language model (LLM) predicts text. Given instructions and context, it can write and edit code impressively well — but it has no memory between sessions. Every new conversation starts from zero unless you supply the context again. Most of what aiflow does is solving exactly this “context problem” in a structured, repeatable way.

What is a context window?

The context window is the model’s working memory: everything it can “see” right now — your instructions, the conversation so far, file contents, tool output. It is large but finite, and every token in it costs money and attention. Two consequences aiflow designs around:

• Don’t fill it with noise. Reading whole files “just in case” wastes the window. aiflow routes questions through a code graph (graphify) and a semantic index (cocoindex-code) so the agent loads only the few relevant chunks.
• Don’t lose what matters. When a session ends (or the window fills up and gets compacted), untracked knowledge dies. aiflow persists it: tasks in Beads, durable facts in memory files, decisions in the issue tracker.

What is Claude Code?

Claude Code is Anthropic’s coding agent for the terminal and IDE. With your permission it reads files, runs commands, edits code, and uses tools. It is configured per project through plain files — CLAUDE.md (the rules every agent follows), .claude/settings.json (permissions + hooks), .claude/agents/ and .claude/commands/. aiflow generates and maintains all of these for you.

What is an agent?

An agent is a focused AI worker with a role and a system prompt: aiflow ships an implementer (senior engineer), a reviewer (architect + quality gate), a tester, an architect, a planner, plus on-demand checkers (security, accessibility, modernisation, …). Each is a small markdown file you can read and edit — the prompt is the configuration. See Agents for what each one does and watches for.

What is memory (in aiflow)?

Three layers, each for a different kind of knowledge:

Layer	Holds	Example
Beads (bd)	tasks, dependencies, status, decisions	“implement order endpoint — in progress, claimed by you”
Memory files (.claude/memory/)	durable prose facts	project aim, dev environment, gotchas
Code indexes (graphify + cocoindex-code)	the structure and meaning of your code	“who calls parseToken?”, “where is retry logic?”

Each session starts by reading these — so the agent begins informed instead of re-discovering your project every time.

What is a skill / slash command?

A reusable instruction you trigger with /name — e.g. /implement (build one task end-to-end), /review-ac (run the review gate), /a11y-check (WCAG audit). They live in .claude/commands/ as markdown files.

What is a hook?

A script the environment runs automatically on events: after the AI edits a file (aiflow auto-formats it), at session start (aiflow pulls the shared issue database), before a git push (aiflow enforces the branching rules). Hooks are how rules stay enforced even when nobody thinks about them.

What is MCP?

The Model Context Protocol is a standard for plugging external tools into the agent: your git host’s issues and PRs, the filesystem, the code graph, live library documentation (context7). aiflow generates the MCP configuration (.mcp.json) from your answers at aiflow init.

What is the Ralph loop?

An autonomy pattern: the agent iterates on a task — implement, verify, fix — until it is COMPLETE or BLOCKED, without you babysitting each step. The implementer decides automatically (from its pre-analysis) whether a task warrants it; you can also request it manually per issue.

Token? Cost?

Models bill by tokens (roughly: pieces of words) — everything the model reads and writes counts. aiflow reduces cost with terse output (caveman), CLI-output filtering (rtk), graph/RAG retrieval, and optional routing of easy steps to cheap or local models. Be aware: the quality rules deliberately spend tokens on tests and reviews — the net saving comes from not having to re-prompt and rework. Measure with aiflow cost.

Where to go next

• Installation — get the CLI.
• Quick Start — first project in minutes.
• Example project walk-through — every question, every default, first feature.
• Agents — who does what, in detail.