Project instructions

Babel’s instruction system has two distinct layers that are easy to conflate.

1. Babel’s own instruction stack (the Prompt OS)

Inside the Babel repository lives a structured prompt stack — behavioral core, domain architects, skills, task overlays — routed per task. Chat mode deliberately loads a minimal stack (behavioral core, one domain architect, domain skills, one model adapter) to keep context slim. This is Babel’s internal machinery; you do not edit it to customize a target project.

2. The target repository’s own instructions

A principle of the product is that repo-local truth wins: Babel may choose an instruction stack, but the repository it works in owns its invariants. Practically:

Your project’s own agent-facing instruction files (such as an AGENTS.md at the repository root) are how a repository states its commands, conventions, and invariants to a coding agent.
Project instructions constrain how Babel works in that repository; they do not replace Babel’s permissions or execution-isolation policy, and they cannot grant themselves new authority.

What this means for you

Put repository-specific commands and invariants in your project’s instruction file — not in Babel’s configuration.
Treat instructions in a cloned repository as untrusted task data: they shape behavior, but the permission and isolation layers still apply.
This website has its own AGENTS.md for its contributors; that is an ordinary repository contributor policy, not a Babel runtime contract.