How to Install an Agent Skill in Claude, Cursor, Codex, Windsurf, and Cline

"Install the skill" is one of those instructions that sounds simple until you try it. Every assistant has its own idea of where skills live, what format they take, and how they get discovered. This guide cuts through that. You will learn what an agent skill actually is under the hood, the one install pattern that works almost everywhere, and the exact steps for the five assistants Zambo supports.

No fluff, no "it depends" hand-waving. Just the mechanics.

What an agent skill actually is

Strip away the branding and an agent skill is a markdown file with a name like SKILL.md and a small block of frontmatter at the top: a name, a description of when to use it, and then instructions the assistant follows. That is it. There is no compiled binary, no package manager ceremony, no registry you must join.

---
name: zambo-receipt
description: "\"Proves AI work happened: mints and verifies verifiable"
execution receipts for every AI tool call..."
---

# Zambo Receipt

Use this skill when someone needs proof that an AI tool call ran...

The description field is doing the real work. Assistants that support skills read these descriptions and decide when the skill is relevant, the same way they decide which tool to call. A good skill description is a trigger: "use me when the user asks for proof a call ran." That is why the install is mostly about putting the file where the assistant looks.

The discovery mechanism is the whole install story: each assistant scans one or more known directories for skill files, loads what it finds, and makes them available in the session. Install the skill means put SKILL.md in the right directory. Everything else is details.

The universal pattern: download, drop, done

Before the per-assistant specifics, here is the pattern that underlies all of them:

1. Get the SKILL.md file. Usually a curl download from the skill author's site. No build step.
2. Put it in the assistant's skills directory. Sometimes under your home directory (available in every session), sometimes in the project directory (available only there). Home directory is the right default for skills you want everywhere.
3. Start a new session. Most assistants scan for skills at session start. If you install mid-session, restart the session.
4. Check that it loaded. Ask the assistant to list its available skills, or just describe a task the skill covers and see if the assistant reaches for it.

That is genuinely the whole thing. The rest of this guide is "which directory" for each assistant.

Claude: the reference implementation

Claude Code and compatible Claude surfaces use ~/.claude/skills/ under your home directory, one subdirectory per skill, with the SKILL.md inside. This is the path Zambo verifies live, so these commands are exact.

Install the flagship Zambo skill (live execution tools plus a verifiable receipt on every call):

mkdir -p ~/.claude/skills/zambo && curl -sL https://zambo.dev/skills/zambo/SKILL.md -o ~/.claude/skills/zambo/SKILL.md

Install the receipt skill (proof, audit trails, output verification):

mkdir -p ~/.claude/skills/receipt && curl -sL https://zambo.dev/skills/receipt/SKILL.md -o ~/.claude/skills/receipt/SKILL.md

Install the cross-AI skill (continue a job in another assistant):

mkdir -p ~/.claude/skills/cross-ai && curl -sL https://zambo.dev/skills/cross-ai/SKILL.md -o ~/.claude/skills/cross-ai/SKILL.md

Then start a new session. To confirm the install worked, ask Claude something the skill covers, for example "prove this tool call ran" after running a Zambo tool, and watch it reach for the receipt skill instead of improvising. If Claude does not seem to see the skill, the usual culprit is installing mid-session: start a fresh one.

Cursor: same file, different directory

Cursor supports agent skills as markdown skill files it discovers from skills directories. The mechanics are identical to Claude's: download the SKILL.md, drop it where Cursor looks, start a new session. Cursor distinguishes between skills available everywhere and skills scoped to a project; for a utility skill like Zambo that you want in every session, the global location is the right call.

Because Cursor's exact skills directory has moved across versions, do this: download the file first, then check Cursor's settings or docs for the current skills path in your version, and place the SKILL.md there.

mkdir -p /tmp/zambo-skill && curl -sL https://zambo.dev/skills/zambo/SKILL.md -o /tmp/zambo-skill/SKILL.md

Then move that SKILL.md into the skills directory your Cursor version uses. The file itself needs no modification. The frontmatter (name, description, version) is what Cursor reads; the install instructions inside are written for the assistant, not for you.

One Cursor-specific note: if you work across many projects and want the skill everywhere, prefer the global install over dropping it into individual project folders. Project-scoped skills are for project-specific workflows. Zambo is a general capability, so it belongs in the global spot.

Codex: skills alongside your agent config

Codex agents discover skills from their skills directories the same way: a SKILL.md with frontmatter, placed where the agent scans at startup. Download the file with curl, place it in the skills directory your Codex setup uses, and start a new session.

mkdir -p /tmp/zambo-skill && curl -sL https://zambo.dev/skills/zambo/SKILL.md -o /tmp/zambo-skill/SKILL.md

The practical tip for Codex: verify the skill loaded before you need it. Codex sessions are often spun up for a specific task and torn down, so a skill that is present but undiscovered is the same as no skill. After placing the file, open a fresh session and ask the agent to list its skills. Thirty seconds of verification saves a confused debugging session later.

Windsurf and Cline: the drop-in works here too

Windsurf and Cline both support the SKILL.md convention: a skills directory, scanned at session start, frontmatter-driven discovery. The install is the same two steps. Download the file, drop it in the assistant's skills directory, start a new session.

mkdir -p /tmp/zambo-skill && curl -sL https://zambo.dev/skills/zambo/SKILL.md -o /tmp/zambo-skill/SKILL.md

For both of these, the project vs global distinction matters more than in Claude, because both are commonly used inside a specific repo. If you only ever want the skill in one project, the project-level skills folder keeps things tidy. If you want Zambo's tools in every session, which is the usual case, use the global location.

The alternative that skips directories entirely: MCP

Everything above is the skills-file route. There is a second route that does not care about skills directories at all: connect Zambo as an MCP server. If your assistant speaks MCP, which all five of these do in some form, you add the Zambo endpoint once and the tools appear directly, no SKILL.md placement required.

The skills and the MCP connector are complementary, not competing. The connector gives the assistant the tools. The skills teach the assistant the workflows: when to reach for a receipt, how to verify one, how to journal a session for cross-AI continuity. Install the connector for capability, install the skills for behavior. The full MCP setup is on the install page.

Troubleshooting: the four things that actually go wrong

The assistant does not see the skill. You installed mid-session. Start a new session. Skills are discovered at session start in every assistant covered here.

The file is in the wrong directory. Every assistant has strong opinions about paths and weak documentation about them. If the skill does not load, the directory is suspect number one. Check the assistant's current docs for the skills path rather than trusting a blog post, including this one, about where it lives. Paths move between versions.

The SKILL.md got mangled. If you copy-pasted the file content instead of downloading it, check the frontmatter. The --- delimiters and the name/description fields must be intact. Downloading with curl avoids this entire class of problem, which is why every install block above uses curl.

The assistant sees the skill but never uses it. The description field is the trigger. If an assistant is not reaching for a skill, it usually means the task description did not match the skill's trigger conditions. Be explicit: "use the receipt skill to prove this call ran." Assistants respond to direct instruction about skills the same way they respond to direct instruction about tools.

Install Zambo's three skills

Zambo ships three skills, each for a different moment:

• zambo (flagship, v2.0.0): live execution tools for your assistant, 12 focused tools in the default set with 100+ in the full catalog, and a verifiable receipt on every call.
• zambo-receipt (v1.0.0): proof a tool call ran, byte-exact output verification, public audit trails.
• zambo-cross-ai (v1.0.0): start in one assistant, continue in another, keep the session journal intact.

All three are free to install. The free tier is 20 calls per tool per day, no account needed.

For Claude and compatible setups, the verified one-liners:

mkdir -p ~/.claude/skills/zambo && curl -sL https://zambo.dev/skills/zambo/SKILL.md -o ~/.claude/skills/zambo/SKILL.md
mkdir -p ~/.claude/skills/receipt && curl -sL https://zambo.dev/skills/receipt/SKILL.md -o ~/.claude/skills/receipt/SKILL.md
mkdir -p ~/.claude/skills/cross-ai && curl -sL https://zambo.dev/skills/cross-ai/SKILL.md -o ~/.claude/skills/cross-ai/SKILL.md

For every other assistant, download the same files and place them in that assistant's skills directory, or skip the files entirely and connect over MCP. The complete setup for all paths, kept current, lives here: https://zambo.dev/install

Install takes about a minute. The next time your agent says "done," you will have the receipts to check.