input-guardrail-planner

Design input validation guardrails for AI agent systems — risk taxonomy, tripwire configs, evaluation benchmarks

SKILL.md

GitHub

Input Guardrail Planner

When to Use

Use when you know you need input validation guardrails for an AI agent but not yet the right design. Best for turning risk concerns into a concrete guardrail architecture before implementation.

Trigger patterns:

  • "Design guardrails for this agent"
  • "What input filters do we need?"
  • "Plan input validation for the agent system"
  • "How should we handle prompt injection / PII / toxicity?"

What It Does

Designs a guardrail architecture covering 5 risk categories (topic drift, toxicity, PII, injection, adversarial prompts) with:

  • Risk taxonomy grounded in the actual agent architecture
  • Tripwire-pattern guardrail configurations (YAML-driven)
  • Precision/recall-calibrated thresholds
  • Runtime approval gates (human-in-the-loop) for high-consequence/irreversible actions — the dynamic complement to static disallowedTools
  • Evaluation benchmark design for each guardrail
  • Implementation task list with TDD rhythm

Invocation

/input-guardrail-planner

Or via meta-router:

/meta-plan "design input guardrails for [agent description]"

Input

Provide:

  1. Agent purpose and capabilities
  2. Target audience and trust level
  3. Data sensitivity classification
  4. Any existing guardrails or filters
  5. Compliance requirements (if applicable)

Output

  • Risk Assessment table (5 categories)
  • YAML guardrail configurations with threshold justification
  • Execution architecture (parallel/sequential)
  • Evaluation benchmarks per guardrail
  • Assumption Register (VERIFIED/REASONABLE/FRAGILE)
  • Implementation tasks

Companion Skills

  • output-guardrail-critic — review guardrail configurations for coverage gaps after design
  • proposal-critic — review the full guardrail design for feasibility and completeness
  • security-threat-model-planner — if the agent needs a broader threat model first