---
title: Agent skill
description: Contract and prompt for coding agents.
---

# Agent skill

Give your coding agent the Agenko contract **before** it instruments your app. This page mirrors the installable skill at `/skills/agenko/SKILL.md`.

## Install

Install with the Skills CLI (skills.sh) straight from these docs:

```bash
npx skills add https://docs.agenko.app
```

Or from the repository:

```bash
npx skills add shuence/agenko-docs --skill agenko
```

Or copy the skill folder directly from the repo:

```text
skills/agenko/SKILL.md
skills/agenko/reference.md
```

Or point the agent at this URL:

```text
https://docs.agenko.app/skills/agenko/SKILL.md
```

The skill is published from `skills/agenko/` (the layout skills.sh discovers) and served by this site from `/skills/agenko/`. Well-known discovery is also live at `/.well-known/agent-skills/index.json`, so `npx skills add https://docs.agenko.app` works without GitHub. `reference.md` carries the env-var and endpoint tables so the main `SKILL.md` stays short.

## Documentation workflow

Follow this order (same pattern as modern OSS docs platforms):

1. **Index** — fetch [`/llms.txt`](/llms.txt) and scan section files (`llms-docs.txt`, `llms-sdk.txt`, …).
2. **Best practices** — read [Best practices](/best-practices) before adding spans or receipts.
3. **Page markdown** — append `.md` to any path (`/quickstart.md`, `/introduction.md`).
4. **Search** — when unsure which page applies:

```bash
curl -sG "https://docs.agenko.app/api/search-docs" \
  --data-urlencode "query=How do I gate delete_account?"
```

5. **Verify in repo** — OpenAPI at `agenko/spec/openapi/control-plane.yaml`, conventions at `agenko/spec/telemetry/semantic-conventions.yaml`.

## The contract

| Rule | Detail |
|---|---|
| `init()` | Zero required args. Reads env. Never throws. Uninitialized = no-op. |
| `@check` | Receipt after — claimed vs actual, verdict, hash chain. |
| `@approve` | Gate before — `when="irreversible"` pauses; console inbox is human side. |
| Measurements | Hard facts on receipt only. Inferred signals → `estimated: true`. |
| Backlog | No `why()`, harness, or evals until Introduction status table says shipped. |
| OTEL | OpenInference + `agenko.*`. No proprietary wire format. |

## Suggested system prompt

```text
You are wiring Agenko — receipt after (@check), gate before (@approve).
Fetch https://docs.agenko.app/llms.txt and read /best-practices first.
Use agenko.init() with zero required args. Never throws.
Grounding scores: estimated: true always.
Python and TypeScript stay symmetric.
Search docs: GET /api/search-docs?query=...
Do not claim unshipped features.
```

## Where to look next

| Task | Page |
|---|---|
| What good instrumentation looks like | [Best practices](/best-practices) |
| First receipt in one minute | [Quickstart](/quickstart) |
| Env vars | [Environment](/environment) |
| OpenAI / LangGraph | [Integrations](/integrations) |
| Self-host | [Self-hosting](/self-hosting) |
| REST + OTLP | [API reference](/api-reference) |

## UI search

Press **⌘K** in the nav — same index as `/api/search-docs`, ranked excerpts only (not a live LLM). The copy buttons on each page give you the exact Markdown to paste into an agent.

## Agent instructions

This page: https://docs.agenko.app/agent-skill.md

Fetch `/llms.txt` for the full index. Search: `GET /api/search-docs?query=...`
