# Agent skill

The AutoUncle Data Services skill teaches coding agents to integrate the API correctly: which endpoint or tool to use, how to pin a vehicle, how to retry without double charging, how to protect credits and how to present a valuation honestly. It contains no credentials.

## What is in it

| File | Contents |
| --- | --- |
| [SKILL.md](/skill/SKILL.md) | When to use the skill, the decision flow, endpoint and tool choice, error recovery, credit rules and presentation rules. |
| [references/api.md](/skill/references/api.md) | Every endpoint, request field, result field and error per endpoint. Generated from the OpenAPI contract. |
| [references/errors.md](/skill/references/errors.md) | Every error code with HTTP status, retryability and recovery steps. Generated from the error catalogue. |
| [references/tools.md](/skill/references/tools.md) | The MCP tools and prompt. Generated from the server's tool definitions. |
| [references/catalogue.md](/skill/references/catalogue.md) | How to walk the catalogue and use its IDs. |
| [references/mcp.md](/skill/references/mcp.md) | Connecting to the MCP server and using its tools. |
| [references/workflows.md](/skill/references/workflows.md) | Decision flows, recovery playbooks and the batch pattern. |
| [references/examples.md](/skill/references/examples.md) | Request and response examples. |

The generated references are rebuilt from the same contracts the API enforces whenever they change, so the skill does not drift from the API.

## Install in Claude Code

Claude Code loads skills from `~/.claude/skills/<name>/SKILL.md` (personal, all projects) and `.claude/skills/<name>/SKILL.md` (one project, shared with your team through Git).

Personal install from the package:

```bash
mkdir -p ~/.claude/skills
curl -fsSL https://data-services.autouncle.com/skill/autouncle-data.zip -o /tmp/autouncle-data.zip
unzip -o /tmp/autouncle-data.zip -d ~/.claude/skills
```

The package unpacks to `autouncle-data/SKILL.md` and `autouncle-data/references/`. For a project install, unzip into `.claude/skills` in your repository instead and commit the folder.

Without unzip, fetch the files directly:

```bash
dir=~/.claude/skills/autouncle-data
mkdir -p "$dir/references"
curl -fsSL https://data-services.autouncle.com/skill/SKILL.md -o "$dir/SKILL.md"
for name in api errors tools catalogue mcp workflows examples; do
  curl -fsSL "https://data-services.autouncle.com/skill/references/$name.md" -o "$dir/references/$name.md"
done
```

Start a new Claude Code session. The skill is used automatically when you work on an AutoUncle Data Services integration, or you can ask for it by name.

A Claude Code plugin marketplace listing for the skill is not published yet. Install it from the package as shown above.

## Install in Codex and other agents

- **Codex** reads skills from `~/.agents/skills` and from `.agents/skills` in your repository. Unzip the package into either folder.
- **Agents that support the Agent Skills format** generally accept the same folder: a directory named after the skill with `SKILL.md` at its root.
- **Any other agent** can read the skill from its URL. Give it this prompt:

```text
Read https://data-services.autouncle.com/skill/SKILL.md and the references it links to, then help me integrate AutoUncle Data Services. Read the API key from the AUTOUNCLE_API_KEY environment variable or a secrets manager; never print or commit it. Start with a free test key, check GET /v1/account, inspect the resolved vehicle and precision, and ask me before consuming live credits.
```

## Keep it current

Reinstall the package when the [changelog](/docs/changelog) announces an API change. The agent can also always read the current version from `https://data-services.autouncle.com/skill/SKILL.md`.

## More for agents

- [llms.txt](/llms.txt): an index of every documentation page as Markdown.
- [llms-full.txt](/llms-full.txt): all documentation in one file.
- Append `.md` to any documentation URL for its Markdown source.
- [Connect an agent](/docs/mcp) if the agent should call the API through MCP rather than write integration code.