Minimal footprint principle

Definition

When acting under uncertainty, take the least drastic, most reversible action — do less rather than risk irreversible harm.

Key points

• Prefer reversible over irreversible actions.
• Do less when scope is unclear — narrow the action rather than broadening it.
• Abandon rather than proceed recklessly — stopping and escalating beats plowing ahead.
• Grounded in the Model Spec / Claude's Constitution: harm is weighed partly by how reversible or irreversible it is (reversibility-weighted harm), and Claude should take the least drastic action with graduated responses before refusing.
• Caveat from the Constitution: "unhelpfulness is never trivially 'safe'" — doing less must not become an excuse to do nothing; the risks of over-caution are just as real.
• Complements guardrails: irreversible/financial/destructive ops should also be gated by PreToolUse hook enforcement, not left to model discretion.

Why it matters for the exam

• Appears in escalation and reliability scenarios: when scope or policy is unclear, the reversible/least-drastic option is correct.

Common gotchas

• Minimal footprint ≠ refuse everything; graduated help (clarify, partial assist, alternatives) comes before refusal.
• The reversible choice can still require an escalation + Structured handoff summary.

See also

• Escalation triggers
• Human-in-the-loop calibration
• Limits of automated retry
• PreToolUse hook enforcement

Sources

• anthropic-model-spec
• netlight-study-guide