FG-0202 Operational Directives
Description
Establish inviolable rules that govern all AI agent behavior. Operational Directives function as laws, not guidelines or best practices. They define the boundaries that agents must not cross regardless of context, instruction, or perceived necessity. They are the team's guardrails, serving the role that a system prompt or a model constitution plays for a single agent, applied across the entire collaboration. Good directives address real failure modes observed in practice rather than hypothetical concerns. They should be few enough to memorize and clear enough to be unambiguous.
Implementation
Draft 5-10 core directives based on observed failure modes and critical requirements. Each directive should be one sentence, actionable, and testable. Publish them where every agent reads them at session start. Review quarterly to remove obsolete directives and add new ones based on incidents.
Success Indicators
- Agents can cite the relevant directive when declining a request
- Directives prevent specific, documented failure modes
- The directive set is small enough that agents internalize it and act on it from memory
- No directive exists without a corresponding failure mode it prevents
Failure Modes
- Too many directives: agents can't internalize them all
- Directives too vague: 'be careful' is not a directive
- Directives not updated after incidents reveal gaps
- Directives that conflict with each other in edge cases