WorkBuddy isn't an image generator; it's my gateway to a clean asset library via Skills

WorkBuddy's Skill Asset Library Entry Point: Clean Up Your Assets First

What really drags down AI design output is often a dirty asset entry point. Today I saw Uishe's second wave of AI and design highlights for September, packed with Skills. In my understanding, a Skill is a reusable task instruction sheet for AI. Web design, content creation, design systems—it makes your fingers itch to try them. My first reaction was that my local pile of PRDs, screenshots, color values, and design-tokens finally had new garbage to feed.

Previously, I wrote about WorkBuddy messing up tables while organizing assets. Later, after integrating DeepSeek and setting up isolated workspaces, it improved slightly. This time, I want to try something more counter-intuitive: the more complete the Skills, the more they drag aesthetics down to stale latte if no one assigns them seats.

My environment is the WorkBuddy web version, which might not apply to everyone. I opened WorkBuddy, clicked Projects on the left, and created a new Design Skill Transit Station. Don't just open the chat box. First, enter Workspace, select Isolated Mode, meaning it won't mix project files with other tables in the same context. Then click Data Sources and add local folders, Feishu docs, and PRD Markdown. I split the materials into two categories: instructions, prompts, and case links go into the Document Library; images, screenshots, and token files go into the Attachment Library. This step is boring but necessary. Previously, throwing structured fields and unstructured attachments together caused WorkBuddy to treat color values as filenames and stuff case screenshots into the owner field.

Next, click Field Configuration and create a table called Skill Assets. Configure fields like this: Asset Name (Text); Type (Dropdown: Design, Content, Frontend, Standards); Source (Text, e.g., Uishe Collection, garden-skills); Applicable Scenario (Multi-line Text); Dependent Token (Text, fill in design-tokens filename); Status (Dropdown: Unverified, Available, Deprecated); Owner (Member). Do not put unstructured screenshots into this table; only associate asset names in the Attachment Library. I expect it won't at least misread a button screenshot as a button.

Then click Automation, create an intake validation rule: block submission if Type is empty; automatically append an External Skill tag if Source contains GitHub or repo names; prompt for completion if Applicable Scenario lacks input or output definitions. Are these rules ugly? Yes, but useful.

AI thrives on boundaries. Without boundaries, it stirs all pretty words into a mush.

I also clicked Model Integration, selected DeepSeek, and hardcoded the Ask-Plan-Craft three-step process in the system prompt: ask clearly, list the plan, produce output. Many design AIs feel heavy-handed because users immediately shout for a high-end page. In WorkBuddy, I made it read the Skill Assets table first, allowing answers based only on items marked as available. For example, if I ask, "Pick three Skills suitable for internal backend prototypes," it outputs selection reasons, waits for my confirmation, then generates an HTML slide draft. The first time, it mixed stitch-skills' token sync with Baoyu's infographics; the interface was okay, but the colors looked like conference room lights at 3 PM—gray enough to make you want to close the window.

I also set permission policies. Product managers have read-only access to the Skill Assets table, seeing only items with 'Available' status; designers can edit Applicable Scenarios and Dependent Tokens; I retain control over Status and Deprecation permissions. In WorkBuddy's Collaboration section, click Roles and assign accordingly. I divide permissions this way mainly because once a design system is messed up, colors look like a cat walked over them. Bug just jumped onto the desk, tail sweeping across the screen; I almost dragged him from 'Available' to 'Deprecated'.

Daily operations are now fixed. Every Friday afternoon, let WorkBuddy run a health check, listing Skills with empty sources, empty scenarios, or those not updated in over thirty days, and output them to a To Be Cleaned sub-table. Each month, put new Skill collections into a sandbox project with read-only permissions, keeping them away from the main library. For this Uishe collection's garden-skills, baoyu-skills, and stitch-skills, I first put them in the sandbox, picked out three or four related to backend design, and marked them as Unverified. My testing shows that installing all Skills is asking for trouble. Core Skills plus one or two supporting ones usually run faster than carrying around a whole warehouse.

The trickiest pitfall is letting it directly output HTML prototypes. Previously, I thought that as long as the model was connected and given a pile of materials, it could serve up a decent page. Now my thinking has changed. WorkBuddy excels at organizing assets into bounded workbenches, but aesthetic judgment still relies on humans. I later changed it to generate a specification checklist first: are colors from tokens? Are font sizes hierarchical? Do components have states? Is spacing uniform? If it passes, let it export available items as HTML slides. Half the ugly stuff disappeared; for the remaining half, I can still scold and fix it myself.

Just now, it dropped the export link into Feishu. I clicked in and took a look; the spacing was still a bit messy. I picked Bug off the keyboard and continued fixing.