Create an assistant by seperated .md files and configuration

[soprano1203]

Hi all,

is it possible to create an assistant by seperated .md files? I only see that I can create a "simple" assitant with some text instructions but without context, e.g. to seperated .md files for further details how to react.

I have an assistant for travel planning which is seperated in multiple files to act with. So a base file for the assistant with instructions in detail and some more files where the assitant/agent should will refer to execute a given task.

Thank you in advance
Tim

[wellkilo]

The Assistant editor itself stores one instructions string; it does not recursively expand Markdown file references placed in that field. For a multi-file agent, current Jan has a better project-level layout: JAN.md + project skills.

For example:

travel-planner/
├── JAN.md
└── .jan/
    └── agent/
        ├── agent.toml
        └── skills/
            ├── itinerary/
            │   ├── SKILL.md
            │   └── templates/
            │       └── itinerary.md
            ├── visa-check/
            │   └── SKILL.md
            └── packing-list/
                └── SKILL.md

• Put the always-applicable travel-planner identity, constraints, and routing rules in JAN.md.
• Put task-specific procedures in .jan/agent/skills/<name>/SKILL.md. Each skill may keep supporting Markdown/templates/scripts beside its SKILL.md and refer to them by relative path.
• Run Jan with that folder as the project root (or use jan --project /path/to/travel-planner). /init can scaffold the project agent files.

The distinction matters for context size: Jan always loads non-empty JAN.md files from the project root and its ancestors. For skills it initially advertises only the name and frontmatter description, then the model calls skill_read to load the full SKILL.md when needed. That gives you separated files without injecting the entire travel knowledge base into every turn.

If a rule must apply to every request, keep it in JAN.md; do not hide it only in a skill that may not be selected. Conversely, large destination-specific procedures are good skills.

This is implemented in the current project-context loader (JAN.md discovery) and skill catalog (progressive disclosure). The built-in Jan guide also documents the project directory layout.

So: not multiple files inside the legacy/simple Assistant instruction field, but yes through a Jan project with JAN.md and separate skills.