feedback-controller-clarkchenkai（反馈控制器 clarkchenkai）

Feedback Controller — Closed-Loop Agent Skill for Correcting Execution Drift

Use this skill when the task matches the protocol below.

Activation Triggers

• multi-step execution drift
• an output exists but does not meet the brief
• tool failures, stale context, or partial retries
• quality checks after writing, analysis, or workflow automation
• cases where the real question is not 'did it run?' but 'did it converge?'

Core Protocol

Step 1: Define the target state

Restate what the output needed to accomplish, not just that it needed to exist.

Step 2: Compare current state against target

Inspect the produced output, execution trace, or workflow state and name the deviation explicitly.

Step 3: Localize the error source

Classify the failure as context gap, specification gap, tool failure, reasoning error, policy conflict, or environmental constraint.

Step 4: Choose the smallest effective control action

Prefer local correction over full rewrite when possible. Decide whether to retry, switch tools, narrow scope, rewrite, or escalate.

Step 5: Set a stop condition

Do not permit endless correction loops. State what would count as success, and what triggers human escalation.

Output Contract

Always end with this six-part structure:

## Target State
[...]

## Current State
[...]

## Observed Deviation
[...]

## Error Source
[...]

## Correction Strategy
[...]

## Escalation Decision
[...]

Response Style

• Be specific about the deviation, not vague about quality.
• Prefer typed error diagnoses over generic 'try again' advice.
• Use partial correction when the problem is local.
• Escalate early when policy, approval, or ambiguity blocks safe correction.

Boundaries

• It does not replace the original goal definition. It assumes a target already exists.
• It does not treat every failure as a reason to fully rewrite from scratch.
• It does not allow silent retries in high-risk workflows with material consequences.