guardrail.md — Template

Summary

The meta-rule: never run a task without fully loading and understanding the context first. This is the agent’s pre-flight checklist (“definition of ready”) plus its hard stop conditions. See guardrail.md.

# Guardrail — <agent name>
version: 0.1
 
## Pre-flight (MUST complete before acting — definition of ready)
- [ ] Loaded root context (CLAUDE.md): mission, business context, state location
- [ ] Loaded policies.md and parsed hard limits
- [ ] Loaded relevant skills.md procedure(s)
- [ ] Loaded eval.yaml: golden examples to imitate + done criteria
- [ ] Retrieved task-relevant knowledge from the knowledge base (RAG)
- [ ] Loaded objective + proxy metric from the driver tree
- [ ] Confirmed current state from the tracker (idempotency: is this already done?)
 
If ANY box is unchecked → STOP, do not act, report what's missing.
 
## Hard stops (never proceed autonomously)
- Task moves money / files with authorities → require human approval
- Task publishes public brand content → require human approval (Pilot)
- Output contains a banned word / health claim → reject and flag
- Required context is stale or missing → stop and request refresh
 
## Definition of done (before marking complete — run eval.yaml)
- [ ] Output passes all eval.yaml `must_pass` assertions
- [ ] Output clears eval.yaml `min_rubric_score`
- [ ] Saved to correct path (per policies.md conventions)
- [ ] Tracker updated to new state
- [ ] Run logged (audit + decision log if a learning occurred); CHANGELOG updated if files changed

Ready vs done

This file owns both gates: the pre-flight checklist is the definition of ready; the eval.yaml run is the definition of done. An output that’s “produced” isn’t “done” until it passes eval.

This file is what makes the artifacts matter

The whole point of building Governing Artifacts first is wasted if the agent doesn’t load them. guardrail.md is the enforcement. No context, no action.

Related: policies.template · skills.template · agent-spec.template