Expert Teams
📖 Best for: YingClaw client users and business teams with recurring, expertise-heavy workloads (copywriting, investment research, contract review, data analysis) who want to crystallize them into an on-demand role.
📖 Reading time: 7 minutes
📖 In one sentence: Turn a role-based AI assistant into a reusable, installable, publishable asset. An Expert is an Agent with a persona plus a skill set; a Team is a Lead agent with several Members—the Lead decomposes and dispatches, each Member executes in their own persona. Entry point is the Expert Teams item in the YingClaw web console sidebar (route
/expert).
1. Core Capabilities
| Capability | Description |
|---|---|
| Create an Expert | Author a role-based AI assistant: persona + skill set + greeting + workflow |
| Create a Team | Pick a Lead and add Members, each with their own persona; the Lead decomposes and dispatches work |
| Two creation modes | Form wizard (clear fields, for users who know what they want) or Conversational creation (tell the AI your needs, it guides you) |
| AI-guided generation | Built-in expert-manager skill—dialog-driven intake that asks follow-up questions and produces the Expert or Team |
| Skill selector | Pick from four sources in one place: Local / YingYu / Tencent / ModelScope |
| Workflow editor | Spell out "first X, then Y" as structured steps; at runtime injected as a "mandatory Expert workflow" |
| Team collaboration | The Lead stays resident in context; Members are loaded on demand—the Lead reads a Member persona only when dispatching, so context is never bloated |
| Import and export | Import and export ZIP packages for sharing and backup |
| Publish to marketplace | Publish a dialed-in Expert for others to discover and install |
| Version update | One-click upgrade when an installed Expert has a new version |
| Marketplace browsing | Filter by type (Expert/Team) and industry; see downloads, favorites, and version count |
2. Typical Use Cases
- Capture an expert playbook — When your best specialist is leaving, encode their topic → hook → structure → compliance workflow as an Expert so everyone gets the same output later.
- Solo or small-team capacity boost — One person cannot cover investment research, legal, and design. Install a few ready-made Experts to fill the gap at low cost.
- Multi-perspective review — A contract needs both a legal view and a business view. Build a Team; the Lead dispatches the contract to two Members, then consolidates into a single opinion.
- Professional output guaranteed — The team charter enforces "Member output is authoritative"—the Lead only orchestrates and compiles, must not skip dispatch and write it themselves.
- Browse by industry — When requirements are fuzzy, browse the marketplace by industry category and install one to try, much faster than writing a persona from scratch.
- Reusable assets — Once an Expert is dialed in, publish it to the marketplace so teammates install instead of reconfiguring.
3. Interface and Operations
1. Entry points
| Item | Description |
|---|---|
| Main entry | Expert Teams item in the web console sidebar (route /expert) |
| Page title | The list page is titled "Expert" and has three tabs: Marketplace / Installed / Published |
| Direct use | Path /agent?expert=<expert_name>&prompt=<first_question>—use the "Use Expert" button on the detail page to jump straight in |
| Conversational create | Path /agent?skill=expert-manager&prompt=...; trigger phrase: "Guide me through creating an Expert or a Team" |
| Conversational edit | Path /agent?skill=expert-manager&edit=<expert_name>&prompt=...; trigger phrase: "Help me edit Expert {name}, full requirements are in the draft" |
Experts and Teams are currently only available in the web console and conversation page. They are not delivered through WeChat, DingTalk, Feishu, or other IM channels, and there is no separate CLI command.
2. Marketplace filters
- Type filter: All / Expert / Team
- Industry category chips: Product Design / Engineering / Game Worlds / Data & AI / Marketing / Content Creation, etc.
- Sorts: Latest / Popular / Favorites
3. Form wizard fields
| Section | Fields |
|---|---|
| Basic info | Name, display name, occupation, summary, tags (specialties), greeting |
| Type switch | Expert / Team |
| Skill selection | Four source tabs: Local / YingYu / Tencent / ModelScope |
| Team fields (shown when Team is selected) | Lead + Members (dynamic add/remove) + team workflow |
| Persona and workflow | Persona body, workflow steps |
When Team is selected, an extra "team workflow" field appears—the AI can auto-orchestrate phases based on the requirements.
4. Expert directory structure
Each Expert = one directory, at path experts/<name>/, containing:
experts/<name>/
├── .codebuddy-plugin/plugin.json # metadata: name, version, type (agent/team), lead, member list, workflow
├── agents/<name>.md # persona doc (YAML header + Markdown body); injected as system prompt at runtime
├── skills/ # copy of skills selected for this Expert
├── avatars/ # avatar
├── workflow.json # workflow step definitions
└── settings.json # Teams only: points to the Lead
A Team adds two things: under agents/ there will be a Lead persona and a persona for every Member; plugin.json carries an extra teamInfo field listing the members.
5. How a Team collaborates
| Stage | Mechanism |
|---|---|
| Who orchestrates | The Lead. At session start only the Lead persona is injected; Member personas are not in the main session—so context is never bloated |
| How dispatch works | The Lead uses the delegate tool to send sub-tasks to Members. On dispatch, the system reads that Member persona as the sub-task's system prompt—the Member executes in their own persona |
| How output is consolidated | Member output returns to the Lead; the Lead consolidates. Members do not communicate with each other directly. |
| Is there a review vote | No. The chain is only Lead dispatches → Member produces → Lead compiles |
| Parallel or serial | Up to the Lead—independent tasks can run in parallel; workflow stages are usually serial |
The charter locks in five rules: the Lead must personally initiate, no writing for Members, no direct Member-to-Member communication, Member output is authoritative, and the Lead may not be dispatched as a Member to themselves.
6. Cards and detail
Expert card: shows name, display name, occupation, summary, tags. Team cards additionally show a team badge + ·N people. Actions change with state (Install / Installed / Update available).
Detail dialog sections
| Section | Content |
|---|---|
| About | Summary and basic info |
| Try Asking Me | Suggested questions, click to start a chat |
| Workflow | Workflow steps |
| Specialties | Tags |
| Team Members | Teams only; lists the Lead and Members; the Lead carries a special purple "Lead" badge |
| Skills | Skills bound to this Expert |
| Version history | Versions and changelog |
Other dialogs: Update version, Publish, Import.
Do not confuse: the web also has a separate "Team" component set (create team / team card / team detail)—that is the skill-sharing team (with a team ID and permissions), not the Expert Team. The names look similar; the functions are different.
4. How to Use
| Scenario | Steps |
|---|---|
| Install an Expert from the marketplace | Expert Teams → Marketplace tab → pick type/industry → click Install |
| Create an Expert via the form | Expert Teams → Form create → fill the 5 sections → submit |
| Create an Expert via conversation | Expert Teams → Conversational create → describe your needs to the AI → AI asks follow-ups and generates |
| Create a Team | Pick "Team" type → pick the Lead → add Members dynamically → configure team workflow |
| Use an installed Expert | Detail page → "Use Expert" → jump to the conversation page → ask your question |
| Publish an Expert to the marketplace | Detail page → Publish → fill in version and notes → submit |
| Upgrade an installed Expert | Installed tab → find "Update available" → click Upgrade |
5. Tips
- Use Conversational create when the requirements are fuzzy — An Expert is only as good as its persona. When in doubt, describe your needs to the AI—it asks follow-up questions and generates better than filling the form blindly.
- Use kebab-case for naming — Names are lowercase English joined by hyphens; the team Lead file is fixed as
<team_name>-team-lead. Avoid reduplicated syllables or random characters. - Do not title the Lead "Team Lead" — Same goes for "Lead" or "Team Lead". Use an orchestration-style title (Investment Editor / Delivery Director)—it directly shapes the Lead voice and self-image.
- Write personas in a five-part structure — Role definition → Core capabilities → Workflow → Output rules → Behavioral constraints. A generic "you are a
{occupation}" persona produces mediocre output. - Only pick skills from the real list — The skill selector only shows what is usable. Do not invent skill names in a hand-written persona—they will silently fail at runtime.
- Put the process in the workflow, not the persona body — The workflow is injected as a mandatory Expert workflow with state and data passing rules. "First X, then Y" belongs in the workflow, not buried in the persona text.
- Keep teams small and sharp — 2 to 3 Members is enough for most review-style scenarios. Member personas are loaded on demand—the more Members, the heavier the Lead's dispatch and consolidation burden.
- Do not let Members talk to each other directly — The charter explicitly says Members do not communicate directly. If your flow needs "A's output becomes B's input", write it in the workflow and have the Lead relay it.
- Watch dispatch quality, not Lead self-writing — Whether the Lead really dispatches, and whether it writes for Members, is the make-or-break factor. If the Lead ends up doing the work itself, reinforce "Member output is authoritative" in the persona.
- Expert Team is NOT the skill-sharing team — The site also has a "Team" component for skill sharing (with a team ID and permissions), not the Expert Team. Similar names, different functions—do not mix them.