SenseNova-Skills: Agent Skills for Office Work, Installed by Copying Directories

The gap SenseNova-Skills fills between a chat model and a finished office artefact

A language model can describe a slide deck. It cannot, on its own, decide which layout fits a quarterly review, generate the images that go on the slides, and keep the whole thing consistent across twenty pages. SenseNova-Skills exists to close that distance. Each skill in the repository is a directory containing a SKILL.md file that declares triggers, capabilities and an execution flow, following the Agent Skills convention. The model reads that declaration, decides when the skill applies, and calls into it. The target user is not a developer building an agent framework from scratch. It is someone who already has an Agent Skills compatible runtime and wants office output: infographics, presentations, spreadsheet analysis, research memos. The README frames the scope as image generation and visualization, slide-deck generation, Excel data analysis, and deep research, usable standalone or composed into end-to-end workflows. Composition is the interesting part. A single skill that generates an image is a utility. A chain that analyses a spreadsheet, generates charts, and lays them out into a deck is a workflow, and that is what the repository is organised around.

How the skill directories, tiers and the sn_agent_runner.py entry point fit together

The architecture visible in the README is layered. The base layer is sn-image-base, described as Tier 0, which exposes low-level tools through a unified sn_agent_runner.py: text-to-image via sn-image-generate, image editing via sn-image-edit, image recognition via sn-image-recognize, and prompt optimisation via sn-text-optimize. The README states this layer is designed to be called by upper-layer skills rather than by a user directly. Above it sits sn-infographic at Tier 1, which performs automatic prompt-quality scoring, selects from 87 layouts and 66 styles, and runs multi-round generation with VLM review and quality ranking before returning a publication-ready infographic. That division matters because it tells you where to look when something goes wrong. If the output image is wrong, the fault is probably in the Tier 0 call. If the layout is wrong, or the prompt scored badly, the fault is in the Tier 1 orchestration. There is also a diagnostic skill, sn-image-doctor, whose stated job is to validate the environment: check that sn-image-base is installed, check Python dependencies, check required environment variables, and interactively write missing values into .env. That is a sensible piece of design. Skill-based agent setups fail most often on environment drift, not on logic, and a skill whose only purpose is to detect that drift is a reasonable investment.

Installing by copying directories into ~/.openclaw/skills or ~/.hermes/skills

The README recommends letting the agent install the skills itself: hand it the repository URL and ask it to clone and place the skills in the right directory. The manual path is short. Clone with git clone https://github.com/OpenSenseNova/SenseNova-Skills.git --depth=1, create the target directory with mkdir -p ~/.openclaw/skills, then copy the subdirectories with cp -r SenseNova-Skills/skills/* ~/.openclaw/skills/. For hermes-agent the target is ~/.hermes/skills/. Two details in the README are easy to miss and will cost time if you do. First, after installation you may need to manually restart the agent service before the new skills are picked up. Second, the API configuration is region-specific. The international endpoint uses base URL https://token.sensenova.ai/v1, while mainland China uses https://token.sensenova.cn/v1, and the README warns explicitly that the docs page, API key, base URL and model name must all come from the same region. Per-category Python dependencies, API keys and invocation examples live in the section guides, for example docs/sn-image-generate_en.md. There is no package manager step and no version pinning mechanism described, which is a real characteristic of this install model rather than a criticism of it.

Why the install model is also the upgrade model, and what that costs

Because skills are directories copied into a runtime's skills folder, updating them means re-copying directories. The README describes no versioning scheme, no lockfile, and no dependency resolver for the skills themselves. The repository has no releases retrieved, so there is no changelog to diff between two checkouts. In practice this means an upgrade is a manual comparison: you either overwrite the directories and accept whatever changed, or you inspect the diff of each SKILL.md before copying. Teams running several agents across machines will feel this. There is no manifest listing which skill versions are deployed where. The Python dependencies add a second axis, since they are documented per category rather than centrally, so a machine that can run the image skills may still lack what the data analysis path needs. The sn-image-doctor skill mitigates the environment half of this problem, but it is scoped to the image stack and its checks are install, dependencies and environment variables. It is a health check, not an upgrade tool. The licence, MIT, removes most distribution concerns and permits modification, but it does not by itself settle the terms attached to the SenseNova Platform API, which are governed separately by the provider.

Where the documentation thins out and the hosted product takes over

The README devotes substantial space to Raccoon, a hosted office product at office.xiaohuanxiong.com that bundles the latest SenseNova models and what it calls the full Cowork-Skill suite, with a free trial and no setup. It also cites adoption figures for that product. Those numbers describe the hosted service, not the repository, and they should not be read as evidence about the open source code. The more practical consequence is documentation asymmetry. The image and visualization section has a named full guide at docs/sn-image-generate_en.md covering prerequisites, Quick Start, API config and invocation samples, and there is a gallery file at docs/sn-infographic-examples.md. The README's own text for the remaining categories is thinner, and the skills list is truncated in the material available here, so the full inventory of directories under skills/ cannot be confirmed. For a reader deciding whether to adopt, that asymmetry matters: the image path is the best-documented entry point, and the slide, spreadsheet and research paths are described at a higher level. Whether that is a temporary state or a deliberate split between open skills and hosted product is not something the README states.

What SenseNova-Skills is not: the case for a plain SDK and a script

A reasonable alternative is to skip the skill layer entirely and call the SenseNova Platform API from your own code. The difference is where the decision-making lives. With an SDK, your script decides the layout, the prompt, the retry policy and the review step, and you can unit test each of those decisions. With SenseNova-Skills, those decisions are encoded in SKILL.md files and executed by the agent at runtime, which is what lets the same skill be reused across tasks without you writing orchestration code, and also what makes the behaviour harder to pin down in a test. If your workflow is fixed and repetitive, a script is the smaller system. If your workflow varies by request and you want the model to choose between 87 layouts and 66 styles based on the content, the skill layer is doing work you would otherwise have to write. The trade is control for adaptability. There is a second alternative worth naming: any general-purpose agent framework with a plugin format. The reason to prefer this repository specifically is the office domain coverage and the SenseNova model pairing, not the plugin mechanism, which follows the public Agent Skills convention and is therefore not unique to this project.

Who should adopt SenseNova-Skills, and what to check before the first run

Adopt it if you run OpenClaw or hermes-agent, or another Agent Skills compatible runtime, and you want office output without building the orchestration yourself. The install is genuinely short and the MIT licence is permissive. Do not adopt it if you need a runtime-agnostic toolchain, if you cannot hold a SenseNova Platform API key, or if you require per-skill version pinning and a changelog before deploying to production. Before the first run, verify four things in order. Confirm your runtime is listed as compatible and that the skills directory path matches the table, ~/.openclaw/skills/ or ~/.hermes/skills/. Confirm your API key, base URL and model name are all from the same region, since the README treats a mismatch as a failure mode. Run sn-image-doctor to have it check the sn-image-base install, Python dependencies and environment variables, and write missing values into .env. Then restart the agent service, because the README states the skills may not be picked up until you do. If the image path works and you plan to use the slide, spreadsheet or research skills, read the corresponding section guide first rather than assuming the image guide covers them.

Editorial conclusion

Adopt SenseNova-Skills if you already run OpenClaw or hermes-agent, you are willing to hold a SenseNova Platform API key, and you want office capabilities that arrive as directories rather than as a service you operate. Do not adopt it if you need a vendor-neutral toolchain, if you cannot accept that part of the suite is documented mainly through the hosted Raccoon product, or if your runtime is not Agent Skills compatible. Before committing, verify three things: that your API key, base URL and model name all come from the same region, that the Python dependencies listed in the per-category guides are installed, and that your agent actually reloads skills after a restart by checking that the copied directories appear under ~/.openclaw/skills/ or ~/.hermes/skills/.

Official sources

1. Issues
2. License: MIT
3. OpenSenseNova/SenseNova-Skills on GitHub
4. README