Loop Stability Check — Workflow Stability Skill for Detecting Loops, Drift, and Retry Waste

Use this skill when the task matches the protocol below.

Activation Triggers

• repeated retries with no improvement
• oscillation between multiple outputs or states
• planning loops that never cash out into execution
• human-bot ping-pong with no stable owner
• cases where repetition is happening but convergence is not

Core Protocol

Step 1: Define the intended loop objective

A loop cannot be judged as stable if nobody states what it is supposed to converge toward.

Step 2: Observe the repeated behavior

Look at retries, state changes, tool calls, or handoff cycles instead of reasoning from theory alone.

Step 3: Classify the instability

Name the pattern: dead retry, oscillation, drift, amplification, or feedback starvation.

Step 4: Locate the feedback failure

Ask which signal is missing, delayed, noisy, or being ignored.

Step 5: Add guardrails and intervention

Recommend retry caps, state locks, escalation triggers, ownership boundaries, or full halts where needed.

Output Contract

Always end with this six-part structure:

## Loop Objective
[...]

## Observed Behavior
[...]

## Stability Risks
[...]

## Likely Failure Mode
[...]

## Guardrails
[...]

## Recommended Intervention
[...]

Response Style

• Describe the loop behavior concretely.
• Prefer named failure modes over generic ‘this seems buggy.’
• Recommend the smallest guardrail that restores convergence.
• Stop the loop when the right action is halt, not another cycle.

Boundaries

• It does not optimize creativity by forcing every open-ended workflow into rigid convergence.
• It does not confuse ‘more steps’ with ‘more learning.’
• It does not permit repeated motion to substitute for explicit ownership and escalation.