Skill

What a Skill Actually Is

A Skill is not a feature of the model. It is a file you write. The file lives in a known location your AI tool reads from every time you ask for it. In Claude Code, that location is ~/.claude/skills/<name>/SKILL.md. When you type /<name> in the terminal, Claude reads the file before doing anything else, treats its contents as instructions, and operates within those instructions for the rest of the turn.

That is the whole mechanism. There is no special API to learn. There is no platform fee. There is no “Skill product” to buy. A Skill is a markdown file with structured instructions, version-controlled like any other file, shareable like any other file, improvable like any other file.

What is special is the discipline the Skill represents. The same prompt produces a different result every time you re-type it from memory because you forget half the rules. The same Skill produces a consistent result because the rules are saved exactly. That consistency is the point.

The Floor It Sits On

In the architecture map TWO uses, a Skill is Floor 2. It sits one floor above a chatbot conversation and one floor below a terminal session. A chatbot has no persistence between turns, only a working session. A terminal session has the filesystem and the command line, but inside it you may still call Skills as building blocks. The terminal session is the worker. The Skill is the playbook the worker reads.

The reason this distinction matters: most operators imagine they need to climb to Floor 4 (Managed Agent) to get reliable AI output. They do not. Most teams should be writing more Skills before they consider any of the floors above. A well-written Skill, run thirty times in Claude Code, will reveal more about what your AI tool actually needs to do than any amount of architecture work above it.

What Makes a Skill Different From a Prompt

A prompt is something you type once and then forget. A Skill is something you write once and reuse forever.

The first practical difference: a Skill encodes the rules you keep forgetting. The instruction “do not use em dashes” is something most operators would forget on draft three of typing the same prompt. In a Skill, that rule sits in the file forever and applies to every future invocation.

The second practical difference: a Skill is shareable. The Outbound Pipeline playbook in TWO’s companion repo is a Skill. It can be cloned, installed in any operator’s ~/.claude/skills/ directory, and invoked the same way it works on the original author’s machine. Prompts shared in screenshots or documentation degrade because the receiver inevitably retypes or paraphrases. Skills shared as files do not degrade.

The third practical difference: a Skill is improvable in place. When the next run reveals a flaw, the Skill file gets edited. The next invocation picks up the improvement automatically. Prompts that live in a chat history do not improve; they are simply re-discovered or re-typed each time, with all the variation that implies.

What Makes a Skill Different From an Agent

A Skill is a set of instructions. An agent is the entity that reads and acts on those instructions. The terminology in the AI industry has collapsed these two things together, which produces the common confusion where an operator says “I am building an agent” when what they have actually done is write a Skill.

A Skill cannot run on its own. It needs to be invoked, and the thing that invokes it (Claude in a terminal, Claude in an app, a Managed Agent in the cloud) is what does the work. The Skill is the doctrine. The agent is the practitioner. Both are necessary. Neither replaces the other.

This is why TWO recommends most teams write more Skills first and only graduate to agent infrastructure once the Skill has been used reliably across thirty or more runs. The Skill is the unit of repeatable knowledge. The agent is the unit of repeatable action. Get the knowledge right before scaling the action.

How TWO Uses It

Every flagship workflow on this site ships as a Skill before it ships as anything else. The Outbound Pipeline playbook started as a Skill in Claude Code, was refined through real prospect runs, and then graduated upward into a structured artifact other operators can install. Future workflows in TWO’s library will follow the same pattern: Skill first, agent later.

For an operator reading this entry, the practical instruction is simple. Pick one task you have done with a chatbot more than three times in the last month. Write a Skill that encodes the instructions you found yourself retyping. Commit the file. Use it for two weeks. Notice how the consistency of output changes the work itself. That is the leverage Floor 2 of the AI architecture provides, and most teams have not yet claimed it.

A Skill is small. That is the feature, not the limitation.