Where model instructions come from

The system prompt is not a string that keeps growing by appending. Rind separates stable identity, workspace rules, runtime capabilities, and session history, and assembles them from their sources before every request.

flowchart TB
    BASE["Base system prompt + environment"] --> CM["ContextManager"]
    DOC["User / project RIND.md"] --> CM
    GOAL["Goal policy"] --> CM
    TEAM["Agent identity / Team catalog"] --> CM
    SKILL["Skill metadata catalog"] --> CM
    HIST[("Session message projection")] --> CM
    CM --> REQ["Model message list"]

The composition root calls build_system_prompt to create the base prompt; a Team Agent's system.md contributes its identity; and when Goal is enabled, the Goal policy is injected as well. ContextManager reads the user-level and project-level RIND.md, injecting at most 32 KiB from each and recording any truncation or read error; the Skill catalog injects only a metadata index, rather than stuffing every Skill body into each turn's request.

Runtime messages are inserted after the first system message, and recent session messages keep their original order. A Goal checkpoint is transient input for a single continuation, and the goal the user states is marked as data rather than a higher-priority instruction. Every category of injection is visible in context stats/decisions, so "why does the model know this?" no longer has to be guessed.

Code entry points: system prompt, RIND.md loading, context assembly. Verification: prompt regression, RIND.md regression.

Back to the series map