🎛️ Feedback Controller

Closed-Loop Agent Skill for Correcting Execution Drift

Observe deviation. Correct fast.

English · 中文 · Philosophy · 哲学说明 · Install Guide · 安装指南 · Examples · 案例

---

What is Feedback Controller?

An open-source Agent Skill that treats the gap between target state and current output as the real work. Instead of celebrating that something was generated, it inspects deviation, localizes the error source, and chooses the right control action: retry, rewrite, switch tools, narrow scope, or escalate to a human.

Built on Norbert Wiener's control-theory insight that systems improve through feedback loops, not one-shot commands. In AI work, that means execution quality depends on how well the system senses deviation and corrects it before error compounds.

---

Who Is This For?

🤖 Agent Builders Use it after multi-step tasks where outputs can drift from intent.	🛠️ Workflow Operators Use it in automations that should retry selectively, not blindly.
📣 GTM & Ops Teams Use it on outbound writing, reporting, and execution that must stay on-target.	⚠️ High-Risk Reviewers Use it when mistakes need escalation instead of silent self-correction.

---

How It Works

graph TD
    A["Task has a target state"] --> B["Observe current output"]
    B --> C["Measure deviation"]
    C --> D["Localize error source"]
    D --> E["Choose correction path"]
    E --> F["Next loop or escalate"]

    style A fill:#fef3cd,stroke:#d4a843
    style C fill:#e8f4f8,stroke:#5ba4c9
    style E fill:#e8f7ec,stroke:#28a745
    style F fill:#eef2ff,stroke:#4f46e5

Loading

The key move: shape the problem the right way before spending more execution on it.

---

Output Contract

Every valid Feedback Controller response ends with the same six-part control diagnosis:

## Target State
## Current State
## Observed Deviation
## Error Source
## Correction Strategy
## Escalation Decision

---

Use Cases

Sales email drifted from the account brief

The output exists, but it is commercially unsafe.

Compare the draft against the account brief, name the specific drift, and correct only the affected sections.

→ Full walkthrough

Market update used stale sources

The system produced a report, but the loop failed on recency control.

Inspect source recency, name the retrieval failure, and rerun only the evidence-gathering and conclusion steps.

→ Full walkthrough

Support reply became defensive

The failure is not factual. It is relational.

Hold the factual core constant, identify the tone drift, and rewrite only the opening and apology logic.

→ Full walkthrough

Automation kept retrying after a permission error

The system is moving, but not learning.

Stop the loop, classify the failure as a non-retriable permission issue, and route it to the owner who can unblock access.

→ Full walkthrough

---

What Makes It Different

Typical Response	Feedback Controller
Treats output existence as success	Treats deviation from target as the real signal
Retries blindly	Chooses a typed correction path
Rewrites everything	Prefers the smallest effective correction
Hides failure behind confidence	Names the failure mode explicitly
Lets loops run indefinitely	Defines stop conditions and escalation

---

Norbert Wiener to AI

• Feedback matters more than one-shot command quality.
• Control is the process of shrinking deviation, not showing off complexity.
• Useful memory is feedback that changes the next action.
• Humans matter most when they provide high-quality corrective signal.

→ Read the full philosophical foundation

---

Quick Install

curl -fsSL https://raw.githubusercontent.com/clarkchenkai/feedback-controller/main/install/install.sh | bash

git clone https://github.com/clarkchenkai/feedback-controller.git
cp -r feedback-controller/feedback-controller ~/.your-platform/skills/feedback-controller

→ Full installation guide

---

Supported Platforms

Platform	Install Method	Status
Claude Code	~/.claude/skills/feedback-controller	✅
Cursor	Remote rule or local skills folder	✅
OpenAI Codex	~/.codex/skills/feedback-controller or .agents/skills/feedback-controller	✅
Gemini CLI	~/.gemini/skills/feedback-controller	✅
Google Antigravity	~/.gemini/antigravity/skills/feedback-controller	✅
Amp / Goose / Cline	~/.agents/skills/feedback-controller	✅

---

Project Structure

feedback-controller/
├── feedback-controller/
│   ├── SKILL.md
│   ├── agents/
│   │   └── openai.yaml
│   └── references/
├── docs/
│   ├── philosophy.md
│   └── philosophy-zh.md
├── install/
│   ├── install.sh
│   ├── README-install.md
│   └── README-install-zh.md
├── examples/
├── examples-zh/
├── scripts/
│   └── validate-docs.sh
├── README.md
├── README-zh.md
└── LICENSE

---

Contributing

Contributions are welcome, especially when they improve:

• sharper protocols and examples
• stronger mirrored Chinese documentation
• better platform compatibility
• clearer high-risk boundaries

---

Changelog

→ Version history

License

Released under the MIT License.