Terms that show up across the skills catalog, the agents pages, and our posts, defined once so other pages can link here instead of repeating them.
- Skills
- Agents
- Classification
- Connectors
- Context Window
- Extraction
- Formatting
- Large Language Model
- Prompt Caching
- Tokens
- Workflow (LLM)
Skills
A skill is a written procedure an AI can follow. It’s a folder with instructions, examples, and sometimes scripts. It captures how a specific job gets done: the steps, the judgment calls, the format of the output. A skill doesn’t run on its own. Something has to invoke it.
Agents
An agent is the thing that runs, and it decides how. It has a model, a set of tools, and a loop: read the situation, decide, act, check, repeat. No one writes that loop out in advance. The agent picks the next move and the tool to use each time around, and keeps picking until the job is done. Agents pick up skills and use them. One agent can run many skills. A skill is the recipe, the agent is the cook.
Classification
Classification is sorting an input into a category from a known list. Is this email a lead, a vendor invoice, or noise? Is this transaction COGS or opex? The possible answers exist before you start; the work is deciding which one applies. In a business brain, classification is what turns a pile of files into a queryable structure, by putting every item into a type the system already understands.
Connectors
A connector is what lets an LLM reach a system of record, instead of just its own context window. It can be an MCP server (Model Context Protocol, the standard way tools plug into a model) or a direct API integration, and either one gives the model read or write access to a real system: a CRM, a bank account, a codebase, a database. Without a connector, the model only knows what’s in the conversation. With one, it can look something up or make a change in the system that actually holds the data.
Context Window
The context window is how much text a model can look back on and reference while generating new text. Think of it as the model’s working memory, not the much larger set of data it was trained on. A larger context window handles longer, more complex prompts; a smaller one limits how much the model can hold onto across a long task or conversation.
Extraction
Extraction is pulling specific facts out of unstructured material and putting them into fields: reading a contract and returning the counterparty, term length, renewal date, and payment terms; reading a call transcript and returning who committed to what. Classification asks “which bucket.” Extraction asks “what are the values.” Extraction is how a business brain gets built in the first place, by turning documents, calls, and what’s in people’s heads into structured records.
Formatting
Formatting is shaping output so it lands where it needs to go, in the form that place expects: the same content as a markdown note for a vault, JSON for an API, HTML for a CMS, or a table for a brief. Formatting carries no judgment about what’s true, only how it’s presented. It’s the cheapest layer to automate, and the one people most often confuse with the actual thinking.
Large Language Model
A large language model (LLM) is trained on a large amount of text and can generate human-like text, answer questions, summarize information, and carry out a wide range of language tasks. Claude, the AI assistant built by Anthropic, is a large language model that’s been further trained and fine-tuned using RLHF.
Prompt Caching
Prompt caching lets a model resume from a specific prefix it’s already processed instead of reprocessing it from scratch. For any task that repeats the same instructions or reference material across multiple calls, this cuts both processing time and cost.
Tokens
Tokens are the smallest individual units a language model works with, and can correspond to words, subwords, characters, or bytes. For Claude, a token is roughly 3.5 English characters, though the exact count varies by language. Tokens are usually invisible when you’re just chatting with a model, but they’re what you’re actually paying for, and what shows up on the bill once a workflow runs at volume.
Workflow (LLM)
Most of the time, a workflow just means the series of steps your team runs to get something done. This is the narrower version. An LLM workflow runs models and tools along a path someone wrote in advance. Step one calls the model, step two checks the answer, step three writes the record, and the order never changes. That is the whole difference between an LLM workflow and an agent. In a workflow the code chooses the path. In an agent the model chooses it. A workflow is easier to predict and cheaper to debug, and it can only do what its path allows.