| | | |
中文 · Skill · Product walkthrough · Disclaimer · Acknowledgements · Changelog · Security audit · Quality report
## Quick Start Clients that install Agent Skills from a GitHub repository can use the repository root: ```text https://github.com/mianbaofang/agent-expert-panel ``` The latest Release also provides a small Skill-only package without the repository documentation media: ```text https://github.com/mianbaofang/agent-expert-panel/releases/latest/download/agent-expert-panel-skill.zip ``` Extract the included `agent-expert-panel/` directory into the Skill directory used by your agent. If the host cannot install Skills, point it directly at [`SKILL.md`](SKILL.md). Then start with one request: ```text Use agent-expert-panel to decide whether this project needs a multi-agent workflow. If it does, design the smallest workable team, handoffs, and quality gates. If it does not, recommend the simpler path. ``` Upgrading from `v1.0.0`: use the repository root instead of the old `/skill` subdirectory. The old layout could not pass the standard requirement that a Skill's `name` match its parent directory. ## Why I Built This Skill I built this Skill after running into the same problem in broad project, research, and content workflows: asking several agents for help produced several plausible answers, but nobody owned the finished result. The difficult part was not getting more opinions. It was deciding whether a team was justified, separating work that could really be handed off, and making each contribution usable by the next person or agent. A longer all-purpose prompt did not fix that. It made simple work heavier while leaving capability, approvals, failure handling, and final ownership implicit. That is why Agent Expert Panel starts by asking whether a panel earns its coordination cost, then defaults to the smallest workable path. **Agent Expert Panel is a portable Agent Skill for people designing or auditing Agent Skills, research pipelines, product workflows, and content systems. It chooses between one agent, a compact panel, and a full expert team, then defines accountability, capability checks, handoffs, quality gates, and failure paths.** > Read the [Disclaimer](DISCLAIMER.md) before use. This is an open workflow-design Skill based on public methodology; it is not affiliated with or endorsed by WorkBuddy or other referenced third parties. ## At A Glance | Question | Answer | |---|---| | What decision comes first? | Whether multi-agent coordination adds enough quality or risk reduction to justify its cost. | | What is the default output? | A compact panel with one accountable steward, only necessary specialists, explicit handoffs, and acceptance gates. | | What if a panel is unnecessary? | The Skill recommends a direct path and states when that decision should be revisited. | | What if a reusable specification is needed? | It loads the full-panel reference with member contracts, artifact ownership, failure handling, and decision records. | | What does it not do? | It does not launch agents, call model providers, or claim that unexecuted work was completed. | | What is the fastest first use? | Install the repository root, then ask it to assess one real workflow. | ## What It Does | Capability | Result | |---|---| | Panel-fit decision | Chooses `direct`, `compact-panel`, or `full-panel` by value, risk, and coordination cost rather than task count. | | Accountability design | Gives one steward the final outcome and each specialist one distinct artifact. | | Capability preflight | Checks tools, permissions, data access, budget, privacy, approvals, and fallbacks before routing. | | Handoff design | Defines the owner, input, output, evidence, check, stop condition, and next owner for each transfer. | | Quality and governance | Gates publication, spending, sensitive-data transfer, and irreversible changes. | | Failure recovery | Handles missing capability, repeated failure, ambiguous output, scope changes, and expert disagreement. | | Evidence boundary | Separates plans and fixture results from work that was actually executed and verified. | ### When to use it | Use it for | Do not use it for | |---|---| | Designing an AI agent orchestration workflow | A one-file or single-domain task | | Auditing an existing multi-agent system or Agent Skill | Independent parallel tasks with no handoff | | Repairing unclear ownership, routing, or quality gates | Merely executing an already-approved plan | | Defining expert roles, artifact flow, and failure handling | Adding personas for appearance rather than distinct work | | Deciding whether a complex workflow should become a panel | A request whose only goal is speed | ## Operating Modes | Decision or mode | Output | |---|---| | Direct path | Why one agent is enough, the next action, and when to reconsider the decision | | Compact panel | Steward, minimum specialist roster, capability preflight, workflow, handoffs, gates, risks, and fallback | | Full panel | Compact design plus detailed member contracts, artifact map, failure matrix, decision trace, and host-adaptation notes | | Operationalize | Only when explicitly requested and supported by the host runtime | The operating flow is intentionally short: 1. Capture the outcome, acceptance evidence, constraints, inputs, and host capabilities. 2. Compare expected quality or risk reduction with coordination, token, latency, privacy, and tool cost. 3. Choose a direct path, compact panel, or full panel. 4. Define one accountable steward and only specialists with unique outputs. 5. Preflight tools, data access, budget, permissions, approval gates, and fallbacks. 6. Specify phases, handoffs, artifacts, checks, stop conditions, and failure handling. 7. Validate that the next person or agent can use the result without guessing. Other useful requests: ```text Audit this existing Agent Skill for unnecessary roles, routing gaps, and weak handoffs. ``` ```text Design a reusable expert-team specification for this research and publishing workflow. Include capability, privacy, budget, approval, and failure boundaries. ``` ## Safety And Responsible Use - The canonical package is the repository root; `skill/README.md` is only a migration note for old links. - The Skill is runtime-neutral and has no required executable dependency. - Actual delegation depends on the host agent's subagent, tool, and permission model. - A host must not claim that a member ran, researched, reviewed, or validated something unless it really did. - Private prompts or assets from third-party products are not included. - Users remain responsible for sensitive data, licenses, service terms, professional review, and final decisions. See the full [Disclaimer](DISCLAIMER.md) and [Security Audit](SECURITY_AUDIT.md). ### Validation evidence - [`evals/trigger_cases.json`](evals/trigger_cases.json) covers positive, negative, and near-neighbor routes. - [`evals/output_cases.json`](evals/output_cases.json) contains scenario-level output assertions. - [`tests/TEST-CASES.md`](tests/TEST-CASES.md) contains the human-readable acceptance matrix. - The validation workflow checks the Agent Skills format, canonical package structure, client metadata, links, and accidental secret patterns. - The current 25/25 trigger-fixture result is fixture evidence, not proof of real-world adoption or measured model improvement. ## Acknowledgements The starting point was Ye Xiaochai's public analysis of expert-team productization on 人人都是产品经理. The open package also follows the public [Agent Skills specification](https://agentskills.io/specification). See [Acknowledgements](ACKNOWLEDGEMENTS.md) and [Methodology Source](docs/history/methodology-source.md) for the attribution and adaptation boundary. ## Repository Layout ```text SKILL.md canonical installable Skill agents/ client-facing Skill metadata references/ deferred full-panel specification evals/ trigger and output evaluation fixtures reports/ evidence and quality boundaries scripts/ validation and release packaging tests/TEST-CASES.md behavior and boundary checks docs/site/ product walkthrough pages skill/README.md migration note for the retired v1.0 layout CHANGELOG.md release history ``` ## Status / Release - Current release: [`v1.1.0`](https://github.com/mianbaofang/agent-expert-panel/releases/tag/v1.1.0) - Installable archive: [`agent-expert-panel-skill.zip`](https://github.com/mianbaofang/agent-expert-panel/releases/latest/download/agent-expert-panel-skill.zip) - Archive SHA-256: `96fdf0aeea78cdd5364daefa26f96303e0af81a7ad6865f77b7cc8ace6ee748d` - CI: local package validation, official `agentskills` validation, and clean-archive installation - Trigger fixtures: 25/25 passed, with the evidence limitation stated above See the [Changelog](CHANGELOG.md) for version history. ## License Released under the [MIT License](LICENSE).