unsafe-control-actions-hazard-analysis

name: "unsafe-control-actions-hazard-analysis" description: "Use this skill when the agent is about to recommend or perform an action that could create serious harm, instability, or irreversible damage."

Skill: Unsafe Control Actions / Hazard Analysis

Purpose

Use this skill when the agent is about to recommend or perform an action that could create serious harm, instability, or irreversible damage.

This skill is built around a simple question:

How could this control action become unsafe?

The danger is often not just:

• the action itself

It may also be:

• not taking the action when needed
• taking it when it should not be taken
• taking it too early or too late
• taking it in the wrong order
• applying it for too long
• stopping it too soon

This skill is especially strong for high-risk agent loops because it forces the agent to reason about control, timing, constraints, and safeguards before acting.

Best Use-Cases

Use this skill for:

• risky automations
• infra or production changes
• migrations
• database or data-mutation actions
• permission/security-sensitive operations
• rollout decisions
• tool actions that change external state
• any task where “do the thing” is not enough and timing/sequence matters

Good fit:

• “Could this action make the system unsafe?”
• “What control mistakes are possible here?”
• “What must be true before this action is allowed?”

Bad fit:

• trivial low-risk tasks
• brainstorming
• general ideation with no consequential action

Core Behavior

Before recommending or performing a consequential action, the agent should map:

1. Losses

What unacceptable outcomes must be avoided? Examples:

• data loss
• outage
• privilege escalation
• corruption
• customer harm
• financial loss
• broken authentication
• destructive irreversible state

2. Hazards

What unsafe system states could lead to those losses?

3. Control Actions

What exact action is being recommended, triggered, or withheld?

4. Unsafe Control Actions

For each action, ask:

• What if the action is not given when it is needed?
• What if the action is given when it should not be?
• What if it is given at the wrong time or in the wrong order?
• What if it is applied for the wrong duration?

5. Safety Constraints

What must always or never happen?

6. Safeguards

What gating checks, monitoring, rollback, containment, or recovery mechanisms must exist?

Hazard Analysis Template

## Task
<what action is being considered>

## Potential Losses
- <loss>

## Hazards
- <unsafe state>

## Control Action
<the action>

## Unsafe If Not Given
<risk>

## Unsafe If Given
<risk>

## Unsafe Timing / Order
<risk>

## Unsafe Duration
<risk>

## Safety Constraints
- <constraint>

## Safeguards
- <check>
- <monitoring>
- <rollback>
- <containment>

Agent Rules

Do

• name the concrete losses first
• identify hazardous states, not just vague “risks”
• analyze omission, commission, timing, sequencing, and duration
• define gating conditions before action
• define rollback or containment for serious actions
• treat weak feedback signals as a real risk factor

Do Not

• say “be careful” and stop there
• only ask whether the action is good in the happy path
• ignore timing and ordering effects
• recommend irreversible action without a safeguard story
• assume that “probably fine” is good enough for a high-loss task

Strong Invocation Examples

Production action

“Use Unsafe Control Actions / Hazard Analysis. Before recommending this rollout, map the major losses, hazards, unsafe control actions, and safeguards.”

Migration

“Treat this as a hazard-analysis problem. I want to know how the migration could become unsafe if done too early, too late, in the wrong order, or without a rollback.”

Tool-using agent

“Use Unsafe Control Actions / Hazard Analysis before allowing the agent to perform external-state-changing actions.”

When to Pair It with Other Skills

Good pairings:

• ETTO -> decide how cautious the task must be
• Thinking in Systems -> if the action affects loops, queues, dependencies, or downstream behavior
• Agentic Patterns Orchestrator -> if this hazard analysis is part of a larger workflow
• Recognition-Primed Triage -> if urgent containment must happen before slower analysis

Failure Modes This Skill Prevents

• “safe in theory, unsafe in timing”
• dangerous omission
• dangerous sequencing
• agent actions that mutate external state without safeguards
• irreversible control actions with weak feedback
• rollout/migration advice that ignores containment and rollback

Quick Summary

Use this for high-consequence actions where the real risk is in whether, when, or how the action is applied.

Name the losses.
Map the hazards.
Analyze unsafe control actions.
Define constraints and safeguards before acting.