18 terms this community uses constantly. Plain definitions, no assumed knowledge! 📖

For if you've been lurking here and feel like you're missing half the conversation...

---

1️⃣ ICM (Interpreted Context Methodology)

💬 "I set up the ICM workspace once with numbered stages. Claude Code reads the folder, knows exactly what to do at each step, and I only review the output at each handoff."

Jake Van Clief's framework for structuring AI agent workflows using folder-based architecture. ICM treats the filesystem itself as the orchestration layer — workflows are broken into numbered stage folders (e.g. 01_research, 02_drafting), each with a CONTEXT.md file that defines what inputs the agent receives, what it should do, and what output it should produce. Designed primarily for use with Claude Code rather than the Claude.ai chat interface. The methodology is model-agnostic and eliminates the need for complex coordination code by using plain text files and folder structure as the state machine.

---

2️⃣ Agent

The most overloaded word in AI right now. Technically: a Claude session or piece of software that can take actions, not just respond. In this community it gets used loosely — anything from a folder-based specialist you consult manually to a fully autonomous system reading emails and triggering workflows without human input. When someone says "I'm building an agent," it's worth asking which kind they mean.

💬 "My agent reads the leads inbox hourly and drafts responses autonomously" is a very different build from "my specialist is a folder I open when I need it."

---

3️⃣ Agentic

The adjective form of agent — but more specific. An agentic workflow is one where Claude takes a sequence of actions, observes the results, and acts again without waiting for human approval at each step. A single Q&A is not agentic. A system that reads an inbox, decides what to do, drafts a reply, and sends it is. The more agentic your build, the more important your rules.md becomes — autonomous action with bad constraints causes autonomous problems.

💬 "The build is fully agentic — it runs on a schedule, processes the leads, and only pings me when it hits a gap it can't resolve."

---

4️⃣ Folder / Folder-based

The core delivery format of ICM. A folder is a collection of structured markdown files — stages, CONTEXT.md files, and reference material — that Claude Code reads as its operating context. Each stage folder defines a bounded unit of work with explicit inputs and outputs. "Folder-based" distinguishes this approach from one-off prompts or standard chat sessions.

💬 "Drop the folder into a new Claude Code session and it's ready to go — no setup conversation needed."

---

5️⃣ Specialist

An ICM folder built to handle one specific, well-defined job — one domain, one output type, one set of rules. Narrow by design. A specialist that can do everything is not a specialist. The word signals intentional scope restriction, and the best ones enforce it with hard refusals at the edges.

💬 "My specialist only produces risk assessment documents for Irish signage contractors. It refuses anything outside that scope — on purpose."

---

6️⃣ Orchestrator

The specialist that receives the initial request and decides which other specialist handles it. Doesn't do the work itself — routes it. In a multi-folder system, the orchestrator is the front door. Without one, the user has to know which folder to open. With one, they just describe the problem.

💬 "The orchestrator reads the incoming message and decides: is this a lead qualification job or a transaction coordination job?"

---

7️⃣ Handoff

The structured document (usually handoff.md) that passes context from one specialist to the next in a multi-agent system. A good handoff tells the receiving specialist what was done, what decisions were made, what gaps remain, and what to do next.

@Ruby Sparks

' design rule from Week 4 of the competition: "If the Gaps field is empty, you're not looking hard enough."

💬 "When the Lead Qualifier finishes, it writes a handoff so the Property Researcher can pick up without asking the client to repeat themselves."

---

8️⃣ Artifact

A self-contained piece of output Claude generates that can be rendered and used directly — most commonly an interactive HTML/JavaScript tool, but also code files, documents, or SVGs. Artifacts let non-developers ship live interactive tools without any deployment step. If someone shares a claude.ai link and it opens a working calculator or form, that's an artifact.

💬 "My submission is a Claude artifact — click the link, it runs in the browser. No GitHub, no Vercel, no hosting."

---

9️⃣ Context window

The maximum amount of text Claude can hold in working memory at once. Everything counts toward it: your folder files, the conversation history, any documents you paste in. When you hit the limit, Claude starts losing earlier parts of the conversation. Most Claude models have a 200,000 token context window — large, but not unlimited, and easy to burn through on big builds.

💬 "My reference folder is 60,000 tokens on its own. Paste a long document on top of that and I'm burning through the context window before we've even started."

---

🔟 Memory

In this community, memory almost always refers to Claude Projects' persistent file storage — content Claude can access across sessions without you re-uploading it each time. Different from the context window (session-only) and from external memory systems like databases or RAG. When someone says "Claude remembered that from last week," they mean it was stored in the Project files.

💬 "I store the client's full brief in Project memory so I never have to paste it again."

---

1️⃣1️⃣ RAG (Retrieval-Augmented Generation)

A technique where relevant documents or data are pulled and fed into the context window at query time — instead of relying on what Claude was trained on. The model retrieves specific information before generating a response. In ICM builds, a well-structured reference/ folder acts as a lightweight RAG layer.

💬 "My legal specialist doesn't rely on Claude's training data for statute citations — it retrieves the actual text from reference/ before answering."

---

1️⃣2️⃣ MCP (Model Context Protocol)

💬 "I wired MCP into the Gmail inbox so the lead qualifier processes new enquiries automatically — no one has to touch it."

An open standard originated by Anthropic — now maintained more broadly — that lets Claude connect to external tools and live data sources: email inboxes, calendars, databases, APIs. Instead of manually copying content in, Claude uses the protocol to call tools and act on results directly. If RAG is "give Claude documents to read," MCP is "give Claude tools to use." Still early but increasingly showing up in advanced builds in this community.

---

1️⃣3️⃣ System prompt

The hidden instructions given to Claude before any conversation starts — sets the rules, persona, and constraints for the entire session. In Claude Projects, your folder files effectively function as an extended system prompt. The user can't see or override it. Your identity.md and rules.md are your system prompt, just structured and separated into readable pieces.

💬 "The system prompt tells Claude it's a WHS specialist and what to refuse — before the user types a single word."

---

1️⃣4️⃣ Claude Code

Anthropic's agentic coding tool, available as a CLI, desktop app, and IDE extension (VS Code and JetBrains). Claude Code can read, write, and edit files directly on your computer and is the primary environment for ICM-based builds. Not the same as Claude.ai (the chat website) or Claude Projects (saved conversations with uploaded files). If someone in the community shipped a live deployed app, a GitHub repo, or a TypeScript intake site, they almost certainly used Claude Code to build it.

💬 "I used Claude Code to scaffold the whole folder structure and push it to GitHub in one session."

---

1️⃣5️⃣ Persona

Two distinct meanings — frequently confused. (1) Brand persona: the voice and character of a business, used in brand and copywriting work. (2) ICM identity: the character defined in identity.md that Claude inhabits when running a specialist. The second is a design decision about how the model presents itself, not a marketing concept. Knowing which one someone means changes the entire conversation.

💬 "The specialist's persona is a no-nonsense ex-vet tech — that lives in identity.md, not the brand guidelines."

---

1️⃣6️⃣ Hallucination

When Claude states something confidently that is factually wrong or entirely made up. Not a traditional software bug — it's a known behaviour of language models that generate statistically plausible text. In ICM builds, good rules.md design (refuse over fabricate, cite sources, flag uncertainty) is the primary mitigation. Most dangerous in research, legal, and compliance contexts.

💬 "My researcher is set to refuse rather than hallucinate. If it can't verify a court ruling, it says so and stops."

---

1️⃣7️⃣ Refusal (as a design principle)

In most AI contexts, refusal = Claude won't do something. In this community, refusal is something you design in deliberately. A well-built specialist refuses to answer outside its scope, refuses to fabricate when it lacks information, and refuses to proceed without what it needs. Across six weeks of competitions, the entries that won used refusal as a load-bearing mechanism — not a fallback.

💬 "Rule 0 in my folder: if there's no verified brief, refuse to produce output. No exceptions."

---

1️⃣8️⃣ Karpathy method / autoresearch

💬 "I ran the autoresearch loop overnight — 30 iterations against my eval set. Woke up to a skill passing 94% of test cases instead of 71%."

A methodology developed by Andrej Karpathy for autonomous iterative improvement of AI systems. The core loop: define a measurable success metric, write eval files (structured test cases with expected outputs), run the agent against them, track pass rates, have the model analyse its own failures and propose changes, apply those changes, and repeat — overnight, automatically, without human supervision at each step. The bottleneck it solves is feedback: without evals and a metric, you can't tell whether a change made things better or worse. With them, you can run dozens of improvement cycles while you sleep.

---

❓ What's missing?

These are the terms I kept tripping over when I joined. There are more — drop your own definition with an example. I'll update the post and put your name next to it 🙏

---

For the full AI glossary (5,000+ searchable terms),

@Bas Rosario

built one and it's excellent:

https://www.skool.com/cliefnotes/prompt-engineering-and-ai-literacy-platform?p=9223c2bd

3 comments