ccpe-system/agents/README.md

218 lines
3.0 KiB
Markdown

# Agents
## 1. Purpose
This directory stores finalized or near-final Agent artifacts.
Agents are reusable AI work roles.
Some are lightweight Prompt Cards.
Some are durable Agent Specs.
Some are committees or multi-agent systems.
## 2. Directory Structure
Recommended subdirectories:
```text
agents/
├── lite/
├── agent-specs/
└── committees/
```
## 3. Subdirectory Usage
### 3.1 lite/
Use for portable prompt cards.
These are suitable for:
```text
Custom GPT
Gemini Gem
Claude Project instructions
Simple expert assistant prompts
Manual chat use
```
Examples:
```text
zhangliao-red-team.prompt.md
cognitive-imaging-specialist.prompt.md
socratic-questioner.prompt.md
```
### 3.2 agent-specs/
Use for durable Agent Specs.
These are suitable for:
```text
Codex
Claude Code
OpenClaw
Multi-agent workflows
Long-term reusable work roles
Agents that call Skills
Agents with authority and evaluation rules
```
Examples:
```text
cognitive-imaging-specialist.agent.md
knowledge-archivist.agent.md
review-committee-chair.agent.md
```
### 3.3 committees/
Use for groups of agents defined together.
This folder may include:
```text
Committee descriptions
Member lists
Role differentiation maps
Committee-level prompts
Links to Runtime Specs
```
Examples:
```text
modeling-committee.md
review-committee.md
writing-committee.md
```
The actual workflow should usually live in:
```text
runtimes/
```
## 4. Agent vs Prompt Card
Use `lite/` when the artifact is:
```text
Portable
Single-role
Directly human-facing
Easy to paste into chat tools
Not tool-heavy
Not state-heavy
```
Use `agent-specs/` when the artifact is:
```text
Durable
Reusable
Workflow-ready
Skill-calling
Authority-aware
Evaluation-aware
```
## 5. Agent vs Skill
An Agent has a role and responsibility.
A Skill is a reusable capability.
If an Agent contains a repeatable method that other Agents can use, consider extracting that method into:
```text
skills/
```
## 6. Agent vs Model Card
An Agent may use a cognitive model.
The model itself should live in:
```text
model-cards/
```
when it is reusable or intellectually important.
Example:
```text
cognitive-imaging-specialist.agent.md
uses:
model-cards/intermediate/cognitive-imaging-model.md
skills/cognitive/cognitive-imaging.skill.md
```
## 7. Naming Convention
Use lowercase kebab-case.
Recommended patterns:
```text
{name}.prompt.md
{name}.agent.md
{name}-committee.md
```
Examples:
```text
zhangliao-red-team.prompt.md
cognitive-imaging-specialist.agent.md
modeling-committee.md
```
## 8. Agent Metadata
Recommended front matter:
```yaml
---
artifact_type:
name:
agent_id:
author:
version:
created:
updated:
status:
target_platform:
related_models:
related_skills:
related_runtimes:
based_on: CCPE System
---
```
## 9. Status Values
Use:
```text
draft
experimental
active
deprecated
archived
```
## 10. Final Rule
Agents are not just personalities.
A good Agent has a role, objective, boundary, workflow, output standard, and relationship to models and skills.