> ## Documentation Index
> Fetch the complete documentation index at: https://www.narev.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Skills repository structure

> Package reusable flight instructions and select the revision to run.

## Start from the supplied skills

Your first flight can use [narevai/pilot-skills](https://github.com/narevai/pilot-skills) with folder `v1`, which the form fills in automatically. No fork is needed to try it.

To customize the agent, fork that repository, edit a `SKILL.md` instruction, then use your fork’s URL in **GitHub skills repository**. Keep `v1` as the folder unless you move the files. Run the same scenario and compare the result.

## Repository layout

The runner clones your repository, checks out the requested ref if provided, and finds `SKILL.md` files recursively under `<folder>/skills/`. Each containing directory is copied into the selected harness's skill directory, including its supporting files.

```text theme={null}
v1/
└── skills/
    ├── flight-workflow/
    │   ├── SKILL.md
    │   └── references/
    │       └── decision-guide.md
    └── approach-monitoring/
        └── SKILL.md
```

Use distinct skill directory names, even when nesting them: installation copies each containing directory into one destination. Keep supporting files inside their skill directory. The runtime removes the temporary repository checkout after installation.

## Write a useful skill

Begin `SKILL.md` with YAML metadata containing `name` and `description`, followed by Markdown instructions. Describe when the skill applies, what evidence to inspect, how to select the next action, and how to check its result.

```markdown theme={null}
---
name: approach-monitoring
description: Use during approach to review the active procedure and cockpit indications.
---

Read the active checklist section and its matching manual chapter.
Discover the relevant cockpit readings rather than inventing IDs.
Compare current indications with the conditions in the procedure.
After an action, verify the result before marking checklist progress.
```

Link to the canonical flight references instead of copying aircraft procedures into every skill. Keep explicit guidance for missing readings, unsupported controls, and unexpected results.

## Repository inspection

The web preview supports public HTTPS GitHub repositories. It resolves the selected ref for inspection, reads skill names and descriptions, and never executes repository code. Preview limits are 10 MiB for the archive, 100 skill documents, and 64 KiB per skill document.

An inspection failure does not prevent submission. A successful preview does not guarantee a successful flight or pin the queued request to that preview's commit. The runner fetches the configured ref when it starts. Use a full commit SHA for reproducible revision selection.

## Common packaging errors

| Symptom | Check |
| - | - |
| No skills found | The folder must contain a `skills/` directory with at least one `SKILL.md` |
| Ref cannot be resolved | Check the branch, tag, or commit exists and is accessible |
| Supporting reference missing | Put it inside the skill directory that gets copied |
| Skills collide | Give each skill directory a unique name |
| Path rejected | Remove parent-directory (`..`) segments |

See [Harnesses and models](https://www.narev.ai/docs/irl-gym/submit/harnesses.md) for the installation destinations.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.