Runtime Skills

Runtime guidance in Bardo exists to keep local tool use disciplined and evidence-first.

Built-in guidanceLink to built-in-guidance#

The MCP server is opinionated about a few core behaviors:

• read local prep artifacts before continuing play
• prefer committed state over free-form narration
• surface uncertainty instead of bluffing
• keep campaign truth local
• treat explicit user corrections as higher priority than older inferred state

Client behaviorLink to client-behavior#

Different clients expose MCP capabilities differently, but the desired behavior is the same: use init to prepare the workspace, inspect readiness and current state, then call the runtime tool that matches the next real need.

Current workspaces also generate a local Agent Skill under .agents/skills/bardo-runtime/. That skill is the portable, project-level entrypoint for connected clients that support progressive skill loading, while .bardo/docs/ remains the deeper local reference layer.

The agent-facing operating contract is simple:

1. read current state
2. read player intent
3. consult relevant rules
4. resolve outcome
5. simulate consequences
6. narrate result
7. propose state changes
8. commit valid changes

If the player explicitly corrects canon, use user_correction before continuing play.

Good Bardo behavior is not just "answer quickly." It is to read the relevant state, name any gaps, resolve only what is supported, and leave table decisions visible when the evidence is incomplete.

PrecedenceLink to precedence#

Precedence should be understood in this order:

1. explicit user correction
2. preserved source rules text
3. current campaign source files
4. approved committed state
5. recent validated play result
6. inference
7. narration flavor

Narration without a validated commit stays advisory, and inference must not override confirmed facts.

How to think about runtime skillsLink to how-to-think-about-runtime-skills#

Runtime guidance is not a replacement for tools. It is there to keep the client honest about what is grounded, what is inferred, and what still needs human judgment. Prefer a safe partial answer over a confident wrong one.