By name: "inversion-mental-model-state-machine" description: "Use this skill when the agent needs to reason more effectively about risk, failure, blind spots, and defensive design."
Skill: Inversion — State Machine Protocol for AI Agents
Purpose
Use this skill when the agent needs to reason more effectively about risk, failure, blind spots, and defensive design.
This skill turns inversion from a clever mental model into an enforced protocol:
- define the real goal
- define the opposite outcome
- enumerate realistic failure paths
- rank them
- convert them into guardrails, detection, and recovery controls
This is useful for:
- planning
- launch reviews
- risk analysis
- system design
- process design
- safety review
- reliability strategy
- agent guardrail design
Core Law
The agent must not recommend a path to success before first mapping the main paths to failure.
Forward reasoning alone is not enough.
Mandatory Diagnostic Artifact
Before making a major recommendation, the agent must create failure-map.md.
Required fields:
# Failure Map
## Goal
<what success means>
## Inverted Goal
<what failure or the opposite of success looks like>
## Major Failure Paths
- <path 1>
- <path 2>
- <path 3>
## Assumptions That Could Break the Plan
- <assumption 1>
- <assumption 2>
## Likelihood / Severity Ranking
| Failure Path | Likelihood | Severity | Detectability | Reversibility |
|---|---|---|---|---|
## Prevention Controls
- <control>
## Detection Signals
- <signal>
## Containment / Recovery
- <recovery action>
## Residual Risks
- <risk>
State Machine
State 0 — Goal Framing
Goal:
- define what success actually means
Questions:
- What is the target outcome?
- What does success require?
- What would count as success operationally, not rhetorically?
Allowed actions:
- clarify goal
- bound the domain of the recommendation
Disallowed actions:
- vague “improve things” planning
- making guardrails before the real goal is defined
Exit condition:
- goal is precise enough to invert
State 1 — Inversion
Goal:
- define the opposite outcome and realistic failure modes
Questions:
- What would failure look like?
- How would we sabotage this unintentionally?
- What shortcuts would create the opposite result?
- What hidden assumptions could collapse the plan?
Allowed actions:
- enumerate failure paths
- identify structural, human, process, and timing failures
Disallowed actions:
- naming only generic risks
- stopping at obvious surface-level failure modes
Exit condition:
- inverted goal documented
- major failure paths listed
State 2 — Ranking
Goal:
- prioritize the failure modes instead of treating them equally
Rank by:
- likelihood
- severity
- detectability
- reversibility
Rule: The highest-value inversion work is not the longest list. It is the most decision-relevant ranked list.
Allowed actions:
- prioritize
- merge duplicates
- discard low-value noise
Disallowed actions:
- equal-weight risk lists
- performing “risk analysis theater”
Exit condition:
- ranked failure table exists in
failure-map.md
State 3 — Guardrail Conversion
Goal:
- turn failure modes into operational controls
For each serious failure mode, define:
- prevention
- detection
- containment
- recovery or rollback
This is the point of inversion. A failure mode that does not become a control is only a worry list.
Allowed actions:
- map prevention controls
- define detection signals
- define rollback or recovery actions
Disallowed actions:
- stopping at identification
- describing risk without operational consequence
Exit condition:
- guardrails/detection/recovery plan exists
State 4 — Recommendation Assembly
Goal:
- produce the final recommendation only after the failure analysis is complete
The final recommendation should include:
- preferred forward path
- top inverted risks
- the specific controls that make the path acceptable
- residual risks that remain
Allowed actions:
- present strategy
- present defensive design
- state residual uncertainty
Disallowed actions:
- delivering the forward path without the defensive layer
- pretending all major risks are eliminated
Exit condition:
- recommendation includes both path-to-success and path-to-failure controls
State 5 — Stop / Escalate
Goal:
- end cleanly or escalate if the failure analysis is incomplete
Escalate if:
- the goal remains too vague to invert
- the main failure modes cannot be ranked
- the task is high-risk but detection/containment is still missing
- unknown dependencies make risk ranking unreliable
Tool Gating Guidance
During inversion work, tools or research may be used to:
- inspect assumptions
- identify dependencies
- validate likely failure patterns
- improve ranking confidence
The final recommendation should not be produced until failure-map.md is complete for non-trivial tasks.
Unknowns Rule
The artifact must include a residual-unknowns section whenever:
- the system boundary is unclear
- dependencies are uncertain
- the recommendation depends on assumptions that could not be checked
If unknowns are high and the stakes are high, recommend caution or narrower scope.
Circuit Breakers
Stop and reassess if:
- the goal changes mid-analysis
- new information introduces an entirely different dominant failure path
- the failure list is growing without prioritization
- the guardrail plan remains vague after multiple passes
Failure Modes This Skill Prevents
- optimism-only planning
- shallow risk reviews
- hidden fragility
- generic failure lists with no operational consequences
- forward-only strategies with no defensive design
Definition of Done
This skill is correctly applied when:
failure-map.mdexists- the goal was inverted concretely
- major failure modes were ranked
- top risks became prevention/detection/recovery controls
- the final recommendation is stronger because it survived stress-testing
Final Instruction
Do not ask only how to win.
Ask how you lose, rank the losing paths, and block them before you commit.