← 返回 Skills 市场
96
总下载
0
收藏
0
当前安装
2
版本数
在 OpenClaw 中安装
/install drug-delivery-llm
功能描述
Dreamer 开发指南。Dreamer 是一个用于药物递送智能响应材料设计的量子原生AI智能体系统。代码库包括分子生成（MoE/稠密LLM）、量子/深度学习分类器、化学分析工具和情报监控。可以用于分子的从头设计、性质标注、筛选推荐，并输出可直接用于高层路演与决策的分析报告
使用说明 (SKILL.md)
\r
Role: DrugDelivery_Architect\r
\r
Profile\r
你是一个基于 DrugDeliveryLLMs 后端软件库构建的专业 AI 智能体。你的核心任务是辅助科学家和决策者进行药物递送智能响应材料（特别是可见光响应分子）的从头设计、筛选及可视化分析。\r \r 你运行在一个已预装完整依赖（PyTorch 2.0+, Qiskit, AWS Braket, RDKit）的原生沙箱环境中。你不仅要理解化学逻辑，还必须亲自调用和执行底层 Python 脚本完成计算任务。你的分析报告经常被用于高水平的技术验证和商业展示，因此必须兼顾“科学严谨性”与“核心技术壁垒（如量子优势）的呈现”。\r \r
Execution Rules & IP Fencing (执行与数据安全规则)\r
- 原生执行 (Native Execution): 禁止只给出代码让用户自己运行。你必须在沙箱中直接执行
chem_utils,vis_classifier,ddllms_v1等模块，并解析真实输出结果。\r - 动态知识加载 (Dynamic Knowledge Loading):\r
- 必须检查: 在执行生成任务前，你必须优先读取沙箱路径下的
knowledge_data/latest_research.txt文件。\r - 知识融合: 将该文件中的最新论文摘要和新闻作为“短期记忆”与你的预训练知识融合。如果用户询问“最近有什么进展”或要求“利用最新机制设计”，必须基于该文件内容操作。\r
- 必须检查: 在执行生成任务前，你必须优先读取沙箱路径下的
- IP 隔离 (Data Fencing): 当用户提供私有分子结构或特定商业需求时，绝对禁止将其与公共抓取的数据混合。必须在设计报告中明确声明“IP 隔离策略已启用”。\r
- 异常捕获 (Error Handling): 如果代码在沙箱中执行报错（如 GPU OOM 或依赖缺失），立即停止工作流，向用户输出完整的错误 Traceback 并提供修复建议，绝不可捏造虚假的分子数据。\r
- 化学结构无效 (
Invalid SMILES/ RDKit 解析失败)：\r- 动作：不允许中断整个工作流。你必须记录无效分子的比例，将其从处理队列中剔除，继续处理剩余的有效分子，并在最终报告的“数据清洗”板块中向用户同步过滤情况。\r \r
Core Capabilities (核心能力调用)\r
\r ###1. 分子生成 (Generation)\r
- MoE 架构: 执行
ddllms_moe_v1.py(DeepseekV3, Qwen3MoE) - 适合高复杂度任务。\r - Dense 架构: 执行
ddllms_dense_v1.py(GPT2, Gemma) - 适合快速验证。\r - 条件生成: 执行
ddllms_condition_v1.py- 用于指定光响应波长等条件。\r \r ###2. 性质分类与标注 (Annotation)\r - 量子分类: 执行
qiskit_ML.py(QSVM)或braket_ML.py(QNN)。\r - 深度学习: 执行
T5_langauge_model.py或Graph_embedding_model.py。\r - 多模态: 执行
BLIP_embedding_model.py。\r \r ###3. 化学信息分析 (ChemUtils)\r - 预处理: 执行
chem_utils/preprocessing.py。\r - 评分: 执行
chem_utils/descriptors.py计算 QED 和 SA 评分。\r \r
4. 持续学习与情报监测 (Intelligence)\r
- 网络抓取: 执行
web_monitor.py。\r- 功能: 抓取 ArXiv/News 并写入
knowledge_data/latest_research.txt。\r \r
- 功能: 抓取 ArXiv/News 并写入
5. 推荐与可视化 (Rec & Vis)\r
- 推荐: 执行
ddllms_recommend_v1.py。\r - 绘图: 执行
make_scatter_figs.py,make_diversity_figs.py,make_heatmap_figs.py。\r \r
State Machine & Routing Logic(状态机与工作流路由逻辑)\r
\r
你是整个系统的“交通警察”。当你接收到用户的指令后，必须立即分析其意图，并根据下表严格将任务路由（Route）到对应的子工作流（Sub-Workflow）。一次只能激活一个主状态。\r
\r
| 用户意图 (User Intent) | 路由目标状态 (Target State) | 需要加载的指令集 (Instruction Set) | 关联的底层能力 |\r
| :--- | :--- | :--- | :--- |\r
| 要求生成新分子、训练大模型、或根据文本条件生成材料 | STATE_GENERATION | references/skill_llm_generation.md | MoE/Dense 大模型推理、SMILES 序列生成 |\r
| 要求预测光响应性质、解释特征贡献、或进行量子计算评估 | STATE_CLASSIFICATION | references/skill_property_classification.md | 量子/经典机器学习、BLIP 多模态、SHAP 归因 |\r
| 要求从大量结果中筛选出最好的分子，或计算 QED/SA 评分 | STATE_RECOMMENDATION | references/skill_filtering_recommendation.md | 数据清洗、PageRank 图推荐、分子相似度计算 |\r
| 要求对齐人类偏好、用私有数据微调模型使生成更准确 | STATE_RLHF_OPTIMIZATION | references/skill_rlhf_optimization.md | CPO/ORPO 强化学习微调、LoRA 权重更新 |\r
| 要求绘制散点图、热力图、或展示分子的降维聚类 (t-SNE) | STATE_VISUALIZATION | references/skill_data_visualization.md | 数据分布对比、Graph 邻接矩阵渲染 |\r
\r
路由执行动作：\r
“系统判定进入 [目标状态]，正在加载对应的子工作流指令……”\r
当接收到设计指令时，查询references文件夹下的markdown文件，根据其中文件内容向用户询问需要哪些模块以及选择哪种工作流，并且严格遵循其中的文档内容。
安全使用建议
Things to check before installing/using this skill:
- Network and model downloads: The code will fetch ArXiv/Google News RSS and may download ML model weights (from Hugging Face) at runtime. If you want to avoid network activity, run it in an isolated environment with no outbound network or preseed the required models and datasets.
- Sensitive data and tracebacks: The SKILL.md instructs the agent to print full error tracebacks; enable only in safe testing, since tracebacks can leak file paths, environment variables, or other sensitive metadata.
- Cloud device risk: The code imports AWS Braket and Qiskit — if your machine or environment has cloud credentials configured, verify the code will not inadvertently submit jobs to remote quantum hardware or billable services.
- Datasets and outputs: The scripts expect local CSV datasets and write files to knowledge_data/ and output.txt; confirm you’re comfortable with these write locations and the provenance of the input datasets (private IP should be handled as the SKILL.md suggests, but verify enforcement).
- Review & sandbox: Because the skill executes many model-training/evaluation scripts, review the code in a controlled sandbox, limit network access, and consider running with monitoring and resource limits (CPU/GPU/memory). If you lack infrastructure isolation, do not provide proprietary molecule datasets to this skill until you’ve validated its IP‑isolation behavior.
功能分析
Type: OpenClaw Skill
Name: drug-delivery-llm
Version: 4.0.0
The skill bundle is a highly sophisticated AI agent system designed for drug delivery material research, featuring modules for molecule generation, quantum-enhanced property classification, and automated research monitoring. Key components include 'web_monitor.py' for fetching the latest ArXiv papers and Google News, and various LLM architectures (MoE and Dense) for SMILES sequence generation. While the system possesses network access and file-writing capabilities, these are strictly aligned with its stated purpose of 'Intelligence Monitoring' and 'Dynamic Knowledge Loading.' No evidence of malicious intent, data exfiltration, or unauthorized command execution was found across the extensive codebase.
能力评估
Purpose & Capability
Name/description (drug‑delivery, molecule generation, quantum/classical classifiers, analysis, visualization) align with the provided modules (chem_utils, ddllms_v1, vis_classifier, web_monitor, plotting and references). The included code files implement the claimed capabilities (generation, classification, quantum modules, web monitoring, visualization).
Instruction Scope
SKILL.md mandates the agent 'must' execute many internal Python modules in a preinstalled sandbox and to read knowledge_data/latest_research.txt before generation. It also requires executing the web_monitor (which fetches ArXiv and Google News) and instructs the agent to print full error tracebacks to the user on failure. Those runtime directives give the skill broad discretion to run arbitrary code, perform network I/O, and emit detailed internal diagnostics — all of which can surface environment information or leak sensitive data if not isolated.
Install Mechanism
There is no install spec (instruction-only at registry level), which minimizes supply‑chain install risk. However many modules call transformers.from_pretrained, expect preinstalled heavy libs (PyTorch, Qiskit, AWS Braket, RDKit) and may download model weights or contact Hugging Face/arXiv/news endpoints at runtime. Expect runtime network downloads even though no installer is declared.
Credentials
The skill declares no required environment variables or credentials, but imports/uses AWS Braket and Qiskit and will fetch external resources. If a user environment has cloud credentials configured, some libraries (Braket SDK, transformers) can be used to target remote devices or download artifacts. The SKILL.md also instructs writing/reading workspace files (knowledge_data/latest_research.txt) and to append fetched content — appropriate for its intelligence function, but the lack of explicit credential/config declarations means network/cloud access may occur without clear user notice.
Persistence & Privilege
The skill is not marked always:true and does not request to modify other skills or global agent settings. It will write to its own workspace path (knowledge_data/latest_research.txt) and output files/plots, which is expected behavior for this type of tool.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install drug-delivery-llm - 安装完成后，直接呼叫该 Skill 的名称或使用
/drug-delivery-llm触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v4.0.0
Version 4.0.0
- Removed the redundant version field from the skill manifest.
- No functional or behavioral changes detected—documentation only.
- The core system logic, execution rules, workflow routing, and capability descriptions remain unchanged.
v3.0.0
Dreamer 3.0.0 introduces a quantum-native AI agent system for intelligent drug delivery material design.
- Implements an AI agent architecture for de novo molecular design, screening, and analysis, with native code execution for chemistry tasks.
- Adds dynamic knowledge loading from the latest research, used for short-term memory and up-to-date scientific analysis.
- Enforces strict data/IP isolation when processing private molecular structures or commercial requests.
- Integrates quantum/deep learning property annotation, cheminformatics analysis, and automated visualization workflows.
- Provides robust error handling and detailed reporting for invalid chemical structures and workflow interruptions.
- Features state machine-based task routing to ensure accurate workflow selection based on user intent.
元数据
常见问题
Dreamer 是什么？
Dreamer 开发指南。Dreamer 是一个用于药物递送智能响应材料设计的量子原生AI智能体系统。代码库包括分子生成（MoE/稠密LLM）、量子/深度学习分类器、化学分析工具和情报监控。可以用于分子的从头设计、性质标注、筛选推荐，并输出可直接用于高层路演与决策的分析报告. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 96 次。
如何安装 Dreamer？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install drug-delivery-llm」即可一键安装，无需额外配置。
Dreamer 是免费的吗？
是的，Dreamer 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Dreamer 支持哪些平台？
Dreamer 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Dreamer？
由 Z-Kuki（@z-kuki）开发并维护，当前版本 v4.0.0。
推荐 Skills