Agentic Coding with TVBO

Why TVBO ships skills for AI assistants

Large language models (Claude, GPT, open-weight models) are increasingly used as pair-programmers — through Claude Code, Cursor, GitHub Copilot, Aider, Codex CLI, or just a raw API call. None of these tools know what TVBO is until you tell them.

TVBO’s skill system is the way TVBO tells them. It’s a bundle of short, focused behavioral rules and reference material that an AI assistant can pull in on demand when you’re working in a TVBO project or maintaining TVBO itself.

You install the skills once, and from that point on every Claude Code / Cursor session in your project benefits from them — without you having to paste a system prompt every time.

pip install tvbo
tvbo skills install --target claude-code

That’s it. Now your AI assistant knows the three pillars (LinkML datamodel, ontology, Python package), the YAML model spec format, how to pick a backend, and how to interact with the TVBO platform at tvbo.charite.de.

Pick your path

You’ve heard of TVBO but haven’t written a model yet, and you want your AI assistant to teach you as you go.

→ Getting started for beginners

You’re using TVBO regularly, you have your own simulation setup, and you want your AI assistant to stop making basic mistakes about the YAML schema or the backend choices.

→ Installing and tuning the skills · Skill reference

You’re contributing to the tvbo library or maintaining the platform. You want the same skill system to guide your AI assistant through the internals — LinkML codegen, backend templates, test markers.

→ Maintainer guide

What’s in the system

Component	What it does
Canonical skills	The source-of-truth markdown files. Live in `skills/` (maintainer) and `tvbo/skills/canonical/` (user, shipped in the wheel).
Renderer	`tvbo/skills/_render.py`. Produces per-tool adapter files from canonical sources.
`tvbo skills` CLI	Subcommands `sync`, `install`, `uninstall` — see installation.
AGENTS.md	A repository-level file at the root of `tvbo/` that gives any AI assistant a 1-minute briefing. → explainer

How it fits together

canonical skill (skills/<name>/SKILL.md)
        │
        ▼   tvbo skills sync (build-time, maintainer side)
        ├──► .claude/skills/<name>/SKILL.md           (Claude Code)
        ├──► .github/instructions/<name>.instructions.md   (Copilot)
        └──► AGENTS.md skills-index region            (universal fallback)

canonical user skill (tvbo/skills/canonical/<name>/SKILL.md)
        │
        ▼   tvbo skills install --target X (install-time, user side)
        ├──► ~/.claude/skills/tvbo-<name>/SKILL.md    (Claude Code)
        ├──► ~/.cursor/rules/tvbo-<name>.mdc          (Cursor)
        └──► stdout / <out>.md                        (prompt — raw API)

The same renderer runs in both contexts. There is exactly one place to edit a rule.

Supported targets

Target	What it produces	Where
`claude-code`	`<dest>/<name>/SKILL.md`	`~/.claude/skills/` (user) or `./.claude/skills/` (project)
`copilot`	`<dest>/<name>.instructions.md`	`.github/instructions/` (repo-only, maintainer)
`cursor`	`<dest>/<name>.mdc`	`~/.cursor/rules/` or `./.cursor/rules/`
`agents-md`	Updates the skills-index region of `AGENTS.md`	`<cwd>/AGENTS.md`
`prompt`	Concatenated markdown (raw API / local Ollama)	stdout or `--out path`
`all`	All applicable targets	as above

Roadmap

Add adapters for Aider, Continue.dev, OpenCode, Codex CLI when those conventions stabilise.
A --check mode in CI to prevent canonical-vs-adapter drift (already available; not yet wired into CI).
A tvbo skills new <name> scaffolder.