技能升级器：使用决策理论v5和RAG将任何技能升级为元技能

skill-upgrader by parcadei/continuous-claude-v3

211 周安装量

3,600 GitHub Stars

GitHub

安装命令

npx skills add https://github.com/parcadei/continuous-claude-v3 --skill skill-upgrader

AI/机器学习自动化知识管理

🇨🇳中文介绍

技能升级器

使用 4 个并行 Ragie 支持的代理，将任何 SKILL.md 升级到决策理论 v5 混合格式的元技能。

使用时机

"将此技能升级到 v5"
"用决策理论形式化此技能"
"为此技能添加 MDP 结构"
"对 X 应用技能升级器"

先决条件

具有索引书籍的 Ragie RAG：

决策理论分区 : LaValle 的《规划算法》、Sutton & Barto 的《强化学习》
模态逻辑分区 : Blackburn 的《模态逻辑》、Huth & Ryan 的《计算机科学中的逻辑》

工作流程

步骤 1：设置会话

SESSION=$(date +%Y%m%d-%H%M%S)-upgrade-{skill_name}
mkdir -p thoughts/skill-builds/${SESSION}

步骤 2：初始化黑板

创建 thoughts/skill-builds/{session}/00-blackboard.md：

# 技能升级：{skill_name}
开始时间：{timestamp}

## 输入技能
{path_to_skill}

## 目标格式
决策理论 v5 混合格式

## 代理发现
（代理将在此下方追加内容）

---

步骤 3：并行启动 4 个代理

使用 Task 工具同时生成所有 4 个代理。每个代理：

广告位招租

在这里展示您的产品或服务

触达数万 AI 开发者，精准高效

联系我们

代理 1：LaValle 规划器

书籍： LaValle 的《规划算法》（决策理论分区） 关注点： 状态、动作、转移

Task(
  subagent_type="general-purpose",
  prompt="""
输入技能：{path}
黑板：thoughts/skill-builds/{session}/00-blackboard.md

你的书籍：Ragie 分区 'decision-theory' 中的 LaValle 的《规划算法》

任务：识别技能中的 MDP 结构。

查询 Ragie：
```bash
uv run python scripts/ragie_query.py -q "MDP state space definition" -p decision-theory
uv run python scripts/ragie_query.py -q "action space sequential decisions" -p decision-theory
uv run python scripts/ragie_query.py -q "POMDP partial observability" -p decision-theory

阅读输入技能并回答：

什么是状态？（阶段、模式、跟踪的信息）
什么是动作？（代理在每个状态下可以做什么）
转移如何工作？（确定性的还是随机的）
这是 POMDP 还是完全可观测的？

写入黑板章节：## 代理 1：状态、动作与转移

格式为带有 LaValle 章节引用的纯英文。""" )

---

## 代理 2：Sutton & Barto 优化器

**书籍：** Sutton & Barto 的《强化学习》（决策理论分区）
**关注点：** 策略、终止、价值
**依赖：** 代理 1

Task( subagent_type="general-purpose", prompt=""" 输入技能：{path} 黑板：thoughts/skill-builds/{session}/00-blackboard.md

你的书籍：Ragie 分区 'decision-theory' 中的 Sutton & Barto 的《强化学习》

等待：首先从黑板读取代理 1 的发现。

任务：设计策略和终止条件。

uv run python scripts/ragie_query.py -q "policy deterministic stochastic" -p decision-theory
uv run python scripts/ragie_query.py -q "episodic termination conditions" -p decision-theory
uv run python scripts/ragie_query.py -q "reward function design" -p decision-theory

使用代理 1 的状态和动作，回答：

策略是什么？（状态 → 动作规则）
何时结束？（终止状态、成功/失败）
奖励是什么？（目标 +，成本 -）
哪些状态具有高/低价值？

写入黑板章节：## 代理 2：策略与价值

格式为带有 Sutton & Barto 章节引用的纯英文。""" )

---

## 代理 3：Blackburn 模态逻辑学家

**书籍：** Blackburn 的《模态逻辑》（模态逻辑分区）
**关注点：** 约束（时态、认知、道义）

Task( subagent_type="general-purpose", prompt=""" 输入技能：{path} 黑板：thoughts/skill-builds/{session}/00-blackboard.md

你的书籍：Ragie 分区 'modal-logic' 中的 Blackburn 的《模态逻辑》

任务：从技能中提取约束。

uv run python scripts/ragie_query.py -q "temporal logic LTL operators" -p modal-logic
uv run python scripts/ragie_query.py -q "epistemic logic knowledge" -p modal-logic
uv run python scripts/ragie_query.py -q "deontic logic obligations" -p modal-logic

阅读输入技能并识别：

时态约束："必须在 Y 之前做 X" → □, ◇, U
认知约束："必须知道 X" → K 运算符
道义约束："必须/禁止/可以" → O, F, P
动态约束："动作导致效果" → [动作]

写入黑板章节：## 代理 3：约束

对于每个约束：

纯英文描述
模态逻辑表示法
为何重要
Blackburn 章节引用 """ )

代理 4：Huth & Ryan 验证器

书籍： Huth & Ryan 的《计算机科学中的逻辑》（模态逻辑分区） 关注点： 验证、安全性、活性 依赖： 代理 1-3

Task( subagent_type="general-purpose", prompt=""" 输入技能：{path} 黑板：thoughts/skill-builds/{session}/00-blackboard.md

你的书籍：Ragie 分区 'modal-logic' 中的 Huth & Ryan 的《计算机科学中的逻辑》

等待：首先从黑板读取代理 1-3 的发现。

任务：验证一致性和完整性。

uv run python scripts/ragie_query.py -q "safety properties verification" -p modal-logic
uv run python scripts/ragie_query.py -q "liveness properties eventually" -p modal-logic
uv run python scripts/ragie_query.py -q "model checking CTL" -p modal-logic

安全性：哪些坏事永远不会发生？□¬(bad)
活性：哪些好事最终会发生？◇(good)
一致性：代理之间是否存在矛盾？
完整性：覆盖范围是否存在空白？

写入黑板章节：## 代理 4：验证

报告每个属性时使用 ✓/✗。总体结论：通过或需要改进 Huth & Ryan 章节引用。""" )

---

## 步骤 4：合成最终技能

所有代理完成后，阅读黑板并创建：

**输出：** `thoughts/skill-builds/{session}/SKILL-upgraded.md`

使用 v5 混合模板：

```yaml
---
name: {original_name}
description: {original_description}
version: 5.1-hybrid
---

# 选项：{name}

## 初始化
[来自原始技能 + 代理 1 的状态分析]

## 观测空间
[来自代理 1 的 POMDP 分析]

## 动作空间
[来自代理 1 的动作]

## 策略
[来自代理 2 的状态→动作规则]

## 终止条件
[来自代理 2 的回合结构]

## Q-启发式
[来自代理 2 的价值指导]

## 约束
[来自代理 3 的模态逻辑]

## 验证
[来自代理 4 的安全性/活性]

用户："将 .claude/skills/implement_plan/SKILL.md 升级到 v5 混合格式"

Claude：
1. 创建会话目录
2. 初始化黑板
3. 并行启动 4 个代理（Task 工具）
4. 等待完成
5. 阅读黑板
6. 合成升级后的技能
7. 报告："升级后的技能位于 thoughts/skill-builds/.../SKILL-upgraded.md"

# 决策理论分区
uv run python scripts/ragie_query.py -q "your question" -p decision-theory

# 模态逻辑分区
uv run python scripts/ragie_query.py -q "your question" -p modal-logic

# 使用重排序以获得更好结果
uv run python scripts/ragie_query.py -q "your question" -p decision-theory --rerank

thoughts/skill-builds/{session}/
├── 00-blackboard.md      # 代理协作
├── SKILL-upgraded.md     # 最终的 v5 混合技能
└── validation-report.md  # 代理 4 的验证报告

2026 年 1 月 22 日

🇺🇸English

Skill Upgrader

Meta-skill that upgrades any SKILL.md to Decision Theory v5 Hybrid format using 4 parallel Ragie-backed agents.

When to Use

"Upgrade this skill to v5"
"Formalize this skill with decision theory"
"Add MDP structure to this skill"
"Apply the skill-upgrader to X"

Prerequisites

Ragie RAG with indexed books:

decision-theory partition : LaValle Planning Algorithms, Sutton & Barto RL
modal-logic partition : Blackburn Modal Logic, Huth & Ryan Logic in CS

Workflow

Step 1: Setup Session

SESSION=$(date +%Y%m%d-%H%M%S)-upgrade-{skill_name}
mkdir -p thoughts/skill-builds/${SESSION}

Step 2: Initialize Blackboard

Create thoughts/skill-builds/{session}/00-blackboard.md:

# Skill Upgrade: {skill_name}
Started: {timestamp}

## Input Skill
{path_to_skill}

## Target Format
Decision Theory v5 Hybrid

## Agent Findings
(Agents append below)

---

Step 3: Launch 4 Agents in Parallel

Use Task tool to spawn all 4 agents simultaneously. Each agent:

Reads the input skill
Queries Ragie for their specific book
Appends findings to the blackboard

Agent 1: LaValle Planner

Book: LaValle's "Planning Algorithms" (decision-theory partition) Focus: States, Actions, Transitions

Task(
  subagent_type="general-purpose",
  prompt="""
INPUT SKILL: {path}
BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: LaValle's "Planning Algorithms" in Ragie partition 'decision-theory'

TASK: Identify MDP structure in the skill.

Query Ragie:
```bash
uv run python scripts/ragie_query.py -q "MDP state space definition" -p decision-theory
uv run python scripts/ragie_query.py -q "action space sequential decisions" -p decision-theory
uv run python scripts/ragie_query.py -q "POMDP partial observability" -p decision-theory

Read the input skill and answer:

What are the STATES? (phases, modes, tracked info)
What are the ACTIONS? (what can agent do in each state)
How do TRANSITIONS work? (deterministic or stochastic)
Is this POMDP or fully observable?

WRITE to blackboard section: ## Agent 1: States, Actions & Transitions

Format as plain English with LaValle chapter citations. """ )

---

## Agent 2: Sutton & Barto Optimizer

**Book:** Sutton & Barto's "Reinforcement Learning" (decision-theory partition)
**Focus:** Policy, Termination, Value
**Depends on:** Agent 1

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Sutton & Barto's "Reinforcement Learning" in Ragie partition 'decision-theory'

WAIT: Read Agent 1's findings from blackboard first.

TASK: Design policy and termination conditions.

Query Ragie:

uv run python scripts/ragie_query.py -q "policy deterministic stochastic" -p decision-theory
uv run python scripts/ragie_query.py -q "episodic termination conditions" -p decision-theory
uv run python scripts/ragie_query.py -q "reward function design" -p decision-theory

Using Agent 1's states and actions, answer:

What's the POLICY? (state → action rules)
When does it END? (terminal states, success/failure)
What are REWARDS? (goals +, costs -)
Which states are HIGH/LOW value?

WRITE to blackboard section: ## Agent 2: Policy & Values

Format as plain English with Sutton & Barto section citations. """ )

---

## Agent 3: Blackburn Modal Logician

**Book:** Blackburn's "Modal Logic" (modal-logic partition)
**Focus:** Constraints (temporal, epistemic, deontic)

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Blackburn's "Modal Logic" in Ragie partition 'modal-logic'

TASK: Extract constraints from the skill.

Query Ragie:

uv run python scripts/ragie_query.py -q "temporal logic LTL operators" -p modal-logic
uv run python scripts/ragie_query.py -q "epistemic logic knowledge" -p modal-logic
uv run python scripts/ragie_query.py -q "deontic logic obligations" -p modal-logic

Read the input skill and identify:

TEMPORAL: "must do X before Y" → □, ◇, U
EPISTEMIC: "must know X" → K operator
DEONTIC: "must/forbidden/may" → O, F, P
DYNAMIC: "action causes effect" → [action]

WRITE to blackboard section: ## Agent 3: Constraints

For each constraint:

Plain English description
Modal logic notation
Why it matters
Blackburn chapter citation """ )

Agent 4: Huth & Ryan Verifier

Book: Huth & Ryan's "Logic in Computer Science" (modal-logic partition) Focus: Validation, Safety, Liveness Depends on: Agents 1-3

Task( subagent_type="general-purpose", prompt=""" INPUT SKILL: {path} BLACKBOARD: thoughts/skill-builds/{session}/00-blackboard.md

YOUR BOOK: Huth & Ryan's "Logic in Computer Science" in Ragie partition 'modal-logic'

WAIT: Read Agents 1-3 findings from blackboard first.

TASK: Verify consistency and completeness.

Query Ragie:

uv run python scripts/ragie_query.py -q "safety properties verification" -p modal-logic
uv run python scripts/ragie_query.py -q "liveness properties eventually" -p modal-logic
uv run python scripts/ragie_query.py -q "model checking CTL" -p modal-logic

Check:

SAFETY: What bad things never happen? □¬(bad)
LIVENESS: What good things eventually happen? ◇(good)
CONSISTENCY: Any contradictions between agents?
COMPLETENESS: Any gaps in coverage?

WRITE to blackboard section: ## Agent 4: Verification

Report with ✓/✗ for each property. Overall verdict: PASS or NEEDS_WORK Huth & Ryan section citations. """ )

---

## Step 4: Synthesize Final Skill

After all agents complete, read the blackboard and create:

**Output:** `thoughts/skill-builds/{session}/SKILL-upgraded.md`

Use v5 Hybrid template:

```yaml
---
name: {original_name}
description: {original_description}
version: 5.1-hybrid
---

# Option: {name}

## Initiation (I)
[From original + Agent 1 state analysis]

## Observation Space (Y)
[From Agent 1 POMDP analysis]

## Action Space (U)
[From Agent 1 actions]

## Policy (pi)
[From Agent 2 state→action rules]

## Termination (beta)
[From Agent 2 episode structure]

## Q-Heuristics
[From Agent 2 value guidance]

## Constraints
[From Agent 3 modal logic]

## Verification
[From Agent 4 safety/liveness]

Example Usage

User: "Upgrade .claude/skills/implement_plan/SKILL.md to v5 Hybrid"

Claude:
1. Creates session directory
2. Initializes blackboard
3. Launches 4 agents in parallel (Task tool)
4. Waits for completion
5. Reads blackboard
6. Synthesizes upgraded skill
7. Reports: "Upgraded skill at thoughts/skill-builds/.../SKILL-upgraded.md"

Ragie Query Reference

# Decision theory partition
uv run python scripts/ragie_query.py -q "your question" -p decision-theory

# Modal logic partition
uv run python scripts/ragie_query.py -q "your question" -p modal-logic

# With reranking for better results
uv run python scripts/ragie_query.py -q "your question" -p decision-theory --rerank

Files Created

After upgrade:

thoughts/skill-builds/{session}/
├── 00-blackboard.md      # Agent collaboration
├── SKILL-upgraded.md     # Final v5 Hybrid skill
└── validation-report.md  # Agent 4 verification

Weekly Installs

211

Repository

parcadei/contin…laude-v3

GitHub Stars

3.6K

First Seen

Jan 22, 2026

Security Audits

Gen Agent Trust HubWarn SocketPass SnykPass

Installed on

opencode203

codex202

gemini-cli200

cursor198

github-copilot197

amp193

AI Elements：基于shadcn/ui的AI原生应用组件库，快速构建对话界面

60,400 周安装

技能升级器：使用决策理论v5和RAG将任何技能升级为元技能

🇨🇳中文介绍

技能升级器

使用时机

先决条件

工作流程

步骤 1：设置会话

步骤 2：初始化黑板

步骤 3：并行启动 4 个代理

相关 Skills

代理 1：LaValle 规划器

代理 4：Huth & Ryan 验证器

使用示例

Ragie 查询参考

创建的文件

🇺🇸English

Skill Upgrader

When to Use

Prerequisites

Workflow

Step 1: Setup Session

Step 2: Initialize Blackboard

Step 3: Launch 4 Agents in Parallel

Agent 1: LaValle Planner

Agent 4: Huth & Ryan Verifier

Example Usage

Ragie Query Reference

Files Created

最新 Skills