---
title: "Instructions"
description: "Standing project context the agent receives on every run"
canonical_url: "https://hoplite.sh/docs/agent/instructions"
markdown_url: "https://hoplite.sh/docs/agent/instructions.md"
---

# Instructions
URL: /docs/agent/instructions
LLM index: /llms.txt
Description: Standing project context the agent receives on every run
Related: /docs/prompting, /docs/agent/skills, /docs/agent

# Instructions

Instructions are project-level context prepended to every run: conventions, architecture notes, things the agent should always or never do. Set them under **Settings → Project → Agent**, alongside the project's model default and [skills](/docs/agent/skills).

Because instructions ride along on every run in the project, they are the right home for anything the agent should never have to be told twice.

## Instructions, skills, or the prompt?

Three places shape the agent's behavior. Pick by how often the guidance applies:

| Put it in | When |
| --- | --- |
| **Instructions** | It applies to every run in the project — coding conventions, architecture notes, standing do's and don'ts |
| **[Skills](/docs/agent/skills)** | It's a reusable workflow the agent should load on demand with its `load_skill` tool, not carry in every run |
| **The thread prompt** | It's specific to this one task — see [Prompting](/docs/prompting) |

## Letting the agent manage its own settings

The agent can inspect and propose changes to project configuration: `project_settings_get` shows how the project's scripts and settings resolve, and `project_settings_update` proposes changes — which always require your approval. See [The agent](/docs/agent) for the full self-configuration model.

## Sitemap

See the full [sitemap](/docs/sitemap.md) for all pages.
Well-known sitemap: [/docs/.well-known/sitemap.md](/docs/.well-known/sitemap.md).