Cerberus System / 刻耳柏洛斯系统

A portable agent skill for inner-part decision councils.
一个可被主流 AI Agent 使用的内在分身议事 Skill。

刻耳柏洛斯系统不是普通利弊分析，也不只是 Codex 专用技能。它是一套可移植的 Agent 工作流：让 AI Agent 引导用户建立 Alpha、Beta、Gamma 三个内在分身，在重要选择前先独立表态，再交叉质询，最后形成带边界、复盘点和异议记录的决议。

Cerberus System is a model-agnostic, agent-portable skill/protocol for meaningful decisions. It can be used by Codex-compatible skill runtimes, or adapted into other agent frameworks as plain instructions.

What It Is

Cerberus helps an agent turn internal conflict into a visible deliberation:

• Alpha: the part that must work, deliver, and remain capable in the world.
• Beta: the part that has a body, emotions, relationships, limits, and daily life.
• Gamma: the part that protects core identity, non-negotiables, dignity, and the person the user refuses to become.

The system does not claim to reveal a perfect “true self.” It builds provisional, inspectable decision lenses from user stories, choices, values, and later corrections.

For the Chinese origin story, product intent, usage guide, and key design turning points, see docs/ORIGIN.zh-CN.md.

Why It Exists

Many AI decision prompts collapse into:

“Here are the pros and cons.”

But hard decisions are rarely just pros and cons. They are internal negotiations.

Should I leave this job?
Should I take this opportunity?
Should I protect stability or chase creative life?
Should I speak up, endure, negotiate, or walk away?

Cerberus gives those inner parts a structured room to speak, challenge one another, revise, and then produce a decision with:

• action;
• boundaries;
• review points;
• stop conditions;
• dissenting views;
• explicit uncertainty.

Repository Structure

cerberus-system/
  SKILL.md
  agents/
    openai.yaml
  references/
    adaptive-interviewing.md
    council-protocol.md
    living-parts.md
    model-and-agent-compatibility.md
    productization.md
    templates.md

The cerberus-system/ folder is packaged as a Codex-compatible Skill, but the core design is framework-neutral. Other agents can use the same Markdown instructions and references as a protocol.

Core Capabilities

• Conversational onboarding instead of form filling.
• Adaptive interviewing for users who struggle to describe themselves.
• Model-agnostic execution across different LLMs and agent frameworks.
• Persistent Cerberus Profiles so users do not need to configure the council every time.
• Identity-neutral Gamma setup that does not assume gender, family structure, or life path.
• Humanized “living parts” so the heads speak from what they protect, fear, and desire.
• Alpha / Beta / Gamma persona card generation.
• Mandatory four-round process:
  1. independent stance;
  2. cross-questioning;
  3. revised stance;
  4. chair synthesis.
• Visible interaction between heads. Cerberus is not valid if it skips cross-questioning and jumps straight to a summary.
• Decision logs and calibration rules.
• Guardrails for privacy, high-stakes decisions, and user agency.

Usage

Codex-Compatible Skill Runtime

Copy cerberus-system/ into your skills directory, for example:

~/.codex/skills/cerberus-system/

Then invoke:

Use $cerberus-system for this decision.

Chinese:

用刻耳柏洛斯系统处理这个决定。

Other Agent Frameworks

If your agent does not support Codex-style skills, use the repository as an instruction package:

1. Load cerberus-system/SKILL.md.
2. Load references only as needed:
   • council-protocol.md for running a full council;
   • adaptive-interviewing.md for users who struggle to express themselves;
   • living-parts.md when the heads feel mechanical;
   • model-and-agent-compatibility.md for weaker models or uncertain environments;
   • productization.md for persistent profiles and reusable UX;
   • templates.md for cards, profiles, briefs, and logs.
3. Follow the user's language.
4. Run all four council rounds. Do not replace the council with a generic summary.

Language

The skill follows the user's language:

• Chinese input: respond in Chinese and call the product 刻耳柏洛斯系统 by default.
• English input: respond in English and call it Cerberus System.
• Mixed input: follow the dominant language.

The folder name remains cerberus-system for compatibility.

Model Behavior

Cerberus does not choose a model by itself. It runs on whichever LLM or agent invokes it.

It adapts to the host model:

• Compact Mode: weaker model, short context, or fast use.
• Standard Mode: default experience.
• Deep Mode: stronger model, richer profile, or complex decision.

Even in compact mode, Cerberus must still show Alpha/Beta/Gamma interaction.

Example Flow

The agent should not begin by asking the user to fill out a long document.

Instead, it asks one question at a time:

Recently, what work decision best represents how you operate?

If the user gives a rich story, the agent extracts patterns.

If the user says “I don’t know,” the agent switches to low-friction choices:

Pick the closest one. At work, when something feels wrong, you usually:
A. Stay quiet and protect yourself.
B. Raise the issue carefully.
C. Directly challenge it.
D. Work around it privately.
E. Start planning an exit.

This makes Cerberus usable for people who are not naturally verbal, reflective, or comfortable with long self-analysis.

Privacy

This repository contains only reusable instructions and generic templates.

It does not include:

• personal onboarding transcripts;
• private persona cards;
• decision logs;
• real user cases;
• private life or work data.

Cerberus should usually work from summaries, not raw private archives.

Reuse Across Sessions

Cerberus supports a persistent profile pattern. After first setup, the agent can save or output a compact Cerberus Profile containing distilled summaries, cards, non-negotiables, and calibration notes.

Future sessions should load that profile and ask only for the current decision, instead of repeating onboarding.

Boundaries

Cerberus is reflective decision support, not a replacement for the user’s judgment.

It should not be treated as professional medical, legal, financial, psychological, or safety advice. For high-stakes decisions, the skill instructs the agent to frame the output as reflection only and recommend appropriate professional support.

Gamma should not become a moral dictator. It can veto only when explicit user-declared non-negotiables are triggered.

Status

This is an early public skill/protocol package. The core protocol is usable, but the system will improve with real-world calibration:

• better onboarding question banks;
• more example councils;
• multilingual refinements;
• compatibility testing across different LLMs and agent frameworks;
• optional scripts for generating local Cerberus workspaces.

License

MIT