SenseNova-Skills – SenseTime's Open-Source Modular AI Office Skills Library

1. What Is SenseNova-Skills

SenseNova-Skills is an open-source modular AI office skills library from the OpenSenseNova (SenseTime) team, designed for Agent runtimes. It breaks down four core office capabilities—infographic generation, PPT creation, Excel data analysis, and deep research—into independent, reusable skill modules. The tool integrates directly into mainstream Agent frameworks such as OpenClaw and hermes-agent, giving AI assistants end-to-end professional office execution without switching between multiple SaaS platforms. Through a tiered skill architecture (Tier 0 foundation + Tier 1 application layer), VLM quality closed-loop mechanisms, and large-file streaming processing, SenseNova-Skills significantly lowers the technical barrier to office automation while maintaining output quality.

Technical positioning and domain: Belongs to AI office automation and Agent skill ecosystem. Its unique positioning is providing an "out-of-the-box" modular skills library rather than a single model or framework. It fills the gap in Agent frameworks for "high-quality office skills," enabling developers to quickly build AI assistants with professional office capabilities.

Development background: Developed by SenseTime's OpenSenseNova team with deep expertise in computer vision, NLP, and multimodal AI. The motivation stems from Agent frameworks handling simple tasks well but lacking end-to-end capability in complex office scenarios (e.g., generating professional infographics, creating standards-compliant PPTs), forcing users to manually operate multiple tools. SenseNova-Skills addresses this through standardized, composable skill modules.

Core value: Solves the "can do but can't do well" problem for AI assistants in office scenarios. Traditional Agents output only text or simple charts; SenseNova-Skills generates high-density infographics (87 layouts, 66 styles), complete PPTX files, cleaned and aggregated data analysis results, and deep research reports with cross-validation. Its innovation treats "skills" as first-class citizens, closing the loop from instruction input to professional output.

Technical characteristics: Uses tiered skill architecture (Tier 0 base capabilities + Tier 1 application logic) where lower-layer capabilities are reused by upper-layer skills, supporting flexible composition. Built-in VLM visual language models perform multi-round quality inspection and scoring to ensure infographic and PPT output meets publishing standards. Supports large-file streaming (10K-row Excel auto-converted to Parquet) and deep research checkpoint resume, balancing performance and reliability.

2. Key Features

• Infographic generation (sn-infographic): One of SenseNova-Skills' standout features. Supports 87 preset layouts and 66 visual styles, automatically transforming complex data, long reports, or text into high-density, highly readable infographics. The core mechanism uses VLM for multi-round quality checks and iterative optimization, ensuring final output meets professional publishing standards in information density, visual hierarchy, and aesthetic style. Especially suited for converting technical whitepapers, industry reports, or operational data into social media assets.

• PPT generation (sn-ppt-entry): Serves as a unified entry point, parsing content from PDF, DOCX, Markdown, plain text, and other formats, automatically routing to two modes: creative mode (16:9 PNG output per slide) and standard mode. The standard mode workflow is rigorous: generate style specifications from input, plan outline, plan asset requirements, perform VLM visual quality checks, then generate HTML and export to PPTX. The entire process ensures unified style and content accuracy.

• Excel data analysis (sn-da-excel-workflow): Provides a complete end-to-end workflow for common office data analysis. Supports simultaneous multi-table reading with automatic format detection; for files exceeding 10K rows, automatically converts to Parquet for processing efficiency. Built-in data cleaning, filtering, and cross-table aggregation let users complete complex analysis via natural language without writing SQL or Python.

• Deep research (sn-deep-research): Breaks research into four phases: planning (auto-generate research plan), multi-dimensional evidence collection (search from multiple sources), synthesis (cross-validate conflicting data), and final report generation. Core advantage is checkpoint resume—intermediate artifacts (plan.json, sub_reports, synthesis.md) persist to report_dir, so interrupted research can resume from the checkpoint, avoiding wasted resources on long-running tasks.

• Vertical search integration: Built-in vertical search engines cover academic (ArXiv, PubMed), developer (GitHub, HuggingFace), Chinese social (Bilibili, Zhihu, Douyin), and English social (Reddit, X, YouTube). AI assistants executing deep research or industry analysis can gather information from multiple trusted sources with cross-validation, significantly improving accuracy and comprehensiveness.

• Resume and brand visual generation (sn-image-resume / sn-image-imitate): sn-image-resume auto-generates polished professional resume images from user info with multiple template styles. sn-image-imitate generates new visual assets matching reference image style—especially useful for teams maintaining brand visual consistency. Both build on Tier 0 text-to-image and image recognition capabilities for WYSIWYG visual content generation.

• Large-file streaming processing (sn-da-large-file-analysis): Designed for Excel files with 10K+ rows. Uses openpyxl read-only mode with iter_rows streaming to avoid loading all data into memory. Automatically converts to Parquet and processes in chunks, enabling efficient large-dataset aggregation analysis even in resource-constrained local environments.

3. How to Use

1. Environment requirements and prerequisites: Requires Python 3.8+ and an Agent runtime such as OpenClaw or hermes-agent. API keys for dependent skills (OpenAI, Claude, or SenseTime proprietary model APIs) must be obtained in advance. Linux or macOS recommended; Windows users should note file path compatibility.

2. Agent auto-install (recommended): The simplest approach. In a deployed Agent (e.g., OpenClaw), send: "Please install SenseNova-Skills from https://github.com/OpenSenseNova/SenseNova-Skills to the skills directory." The Agent clones the repo and copies skills to the correct directory. Restart the Agent service after installation. Suited for users unfamiliar with command line.

3. Manual installation: Clone the repo: git clone https://github.com/OpenSenseNova/SenseNova-Skills.git --depth=1 (--depth=1 reduces clone size). Copy the skills directory per Agent framework: OpenClaw users to ~/.openclaw/skills/, hermes-agent users to ~/.hermes/skills/. Enter each skill directory and configure .env per SKILL.md guidance with required API keys. Restart Agent service to trigger skills via natural language.

4. Key configuration notes: Each skill's SKILL.md is the core config file declaring trigger conditions, required capabilities, execution flow, and environment variables. Modify .env variables like OPENAI_API_KEY, ANTHROPIC_API_KEY for your model APIs. For local deployment, configure SenseTime proprietary or open models (Qwen, DeepSeek).

5. Direct use (zero-config option): Visit the "Office Raccoon" (Raccoon Pro) platform. SenseNova-Skills skills are pre-integrated in Raccoon Pro plans—no API keys or environment setup required. Experience infographic generation, PPT creation, data analysis, and more. Best entry point for non-technical users.

4. Pros and Cons

5. Comparison with Similar Tools

Selection recommendations:

• Teams pursuing end-to-end office automation: If core need is one-stop automation from data analysis to PPT reporting with controllable output quality (infographics, professional PPT), SenseNova-Skills is currently the best choice. Built-in VLM QC and checkpoint resume are differentiators other solutions lack.
• Enterprises deeply integrated with Microsoft ecosystem: If fully on Microsoft 365 with high data security requirements (enterprise compliance), Microsoft Copilot Cowork's zero-config and ecosystem integration is irreplaceable. Note: no explicit output quality checks and no private deployment.
• Technical teams or highly customized scenarios: If strong AI engineering capability and flexible tool composition (RAG, databases, external APIs) are needed, LangChain offers maximum freedom but requires writing office-specific skill logic with higher development cost.

6. Editor's Review

From technical innovation, SenseNova-Skills makes several notable contributions to AI office automation. First, its "tiered skill architecture" is an elegant design—decoupling base capabilities (text-to-image, image recognition) from application logic (infographic generation, PPT creation), enabling developers to compose skills like building blocks—a forward-looking design in the Agent ecosystem. Second, the VLM quality closed-loop is a true differentiator, solving AI content generation's core "can do but can't do well" pain point through multi-round visual QC ensuring publishing-standard output—rare among similar tools. Third, checkpoint resume and large-file streaming reflect deep understanding of real usage scenarios—long-running research and large dataset processing are common yet often overlooked challenges in office automation.

In practical value, SenseNova-Skills' biggest advantage is out-of-the-box end-to-end capability. Users need not switch between SaaS platforms (Canva, PowerPoint, Excel, Perplexity)—complete chains from data analysis to final reporting via natural language. For operations, marketing, and research roles needing rapid high-quality content, efficiency gains are substantial. Open source and private deployment support provide natural advantages in enterprise data security scenarios.

Clear target audience: technical users deploy and customize via SKILL.md for deep integration; non-technical users experience via Raccoon Pro zero-config. This dual-track strategy covers individual developers to enterprise teams.

Future potential: as Agent ecosystems mature, skill libraries as "capability middleware" grow increasingly important. If SenseNova-Skills expands skill scale (project management, contract review, etc.) and establishes community contribution mechanisms, it could become the "standard skills library" for AI office automation. API dependency and Chinese ecosystem refinement remain key optimization directions.

7. Application Scenarios

• Monthly operations report automation: Operations staff clean and aggregate monthly Excel data via sn-da-excel-workflow, supplement industry trend insights via sn-deep-research, then generate unified, accurate reporting PPT via sn-ppt-standard. Entire process without manual Excel or PowerPoint operation, significantly reducing repetitive work.

• Industry research report generation: Input an industry name (e.g., "2026 new energy vehicle market"); SenseNova-Skills auto-executes research planning (sn-research-planning), multi-dimensional evidence collection from academic databases (ArXiv, PubMed), developer platforms (GitHub), and social platforms (Zhihu, Bilibili) (sn-dimension-research), cross-validates conflicting data, and synthesizes Markdown + offline HTML reports with charts (sn-research-synthesis). Suited for consulting and investment research institutions.

• Infographic marketing content production: Marketing teams transform product selling points, technical whitepapers, or operational data into high-density infographics with 87 layouts via sn-infographic, adapted for WeChat, Xiaohongshu, Bilibili, and other platforms. VLM QC ensures visual hierarchy and aesthetic quality without professional designers.

• Resume and brand visual asset generation: Job seekers use sn-image-resume for personalized resume images with multiple template styles. Enterprise teams use sn-image-imitate to generate brand-consistent visual assets for social media and promotional posters.

• Large-file financial data analysis: Finance teams facing 10K-row sales or financial tables use sn-da-large-file-analysis for streaming read and Parquet conversion with aggregation analysis, avoiding memory overflow. Especially suited for annual sales data and customer transaction records.

8. FAQ

Q: Does SenseNova-Skills require a GPU to run?
A: No. SenseNova-Skills is a skill library and doesn't run model inference itself. It depends on external LLM and VLM APIs (OpenAI, Claude, or SenseTime proprietary models). GPU requirements depend on your chosen API service. SenseTime cloud API requires no local GPU; local open-source model deployment requires GPU configuration.

Q: How does VLM quality inspection in the skill library work?
A: VLM QC mainly applies to infographic and PPT generation. During generation, the system calls visual language models (GPT-4V or SenseTime VLM) for multi-round scoring on dimensions including information density, visual hierarchy, color coordination, and text readability. If scores fall below threshold, the system auto-adjusts parameters and regenerates until output meets preset quality standards.

Q: Can I customize skills? Add new data sources or modify generation logic?
A: Yes. SenseNova-Skills is fully open source (Apache 2.0). Each skill's SKILL.md defines trigger conditions, capabilities, and execution flow. Modify SKILL.md to adjust behavior or create new skills based on Tier 0 foundation. Advanced users can modify Python source for deep customization.

Q: What's the difference between Raccoon Pro and the open-source version?
A: Raccoon Pro is SenseTime's commercial version with all SenseNova-Skills pre-integrated—no environment, API key, or Agent runtime setup required. Zero-config web interface suits non-technical users. Open source allows developers to deploy, modify, and extend for private deployment or deep customization. Core skill capabilities are consistent; Raccoon Pro may offer additional commercial support and performance optimization.

Q: Which Agent frameworks does the skill library support?
A: Officially supports OpenClaw and hermes-agent. OpenClaw is SenseTime's Agent framework; hermes-agent is a popular open-source Agent runtime. Following unified SKILL.md specification, integration with any compatible Agent framework is theoretically possible. Community users are experimenting with LangChain and AutoGPT integration.

Q: How to handle API key security?
A: Each skill's .env stores API keys, default ignored by .gitignore and not committed to version control. For enterprise deployment, use environment variables or secret management services (HashiCorp Vault) rather than writing directly to .env. Private deployment can restrict API keys to specific IPs for additional security.

9. Project Links

• GitHub repository: https://github.com/OpenSenseNova/SenseNova-Skills
• OpenClaw framework: https://github.com/OpenSenseNova/OpenClaw