Your AI expert roundtable.
Turn one model's answer into a task-specific panel of Lingtai agents that challenges plans, maps blind spots, and keeps one Executor accountable for delivery.
中文 · English
Roundtable Skill turns a single-agent task into a small expert meeting. The Executor assigns temporary expert roles to available Lingtai agents, collects their reviews, maps disagreements, then decides what to do next.
It works best when one model thinking alone is too narrow: release gates, research briefs, strategy decisions, business plans, product reviews, and other work where missing one angle can change the answer.
First time here? Start with the Quickstart 60-second fit check.
- 🪑 Runs a real roundtable: each Lingtai agent gets a temporary expert role for this task only.
- 🔍 Finds blind spots: agents review the first plan, the evidence, and the Executor's progress from different angles.
- ⚖️ Maps disagreement: conflicting claims, weak evidence, and missing perspectives are made explicit.
- ✅ Keeps one owner: the Executor still owns the final decision, implementation, verification, Git state, and rollback when code is involved.
- 🧯 Handles silent agents: no infinite waiting; record the non-response, try one safe repair, then continue with evidence.
A single agent can sound confident while missing the point.
Roundtable adds structured disagreement. A practitioner may catch what the paper ignores. A skeptic may find the strongest counterexample. A security reviewer may block a risky merge. A finance voice may notice incentives that the product voice missed.
The roles are not permanent identities. They are assigned for the current task and discarded when the task ends.
The same roster can be reshuffled for every problem: one agent may be a skeptic for a research brief, a release reviewer for a deploy gate, and a customer voice for a business plan. The value comes from choosing useful tension, not from locking agents into permanent personas.
The point is not "more agents." The point is the right tension: different roles pull on the same problem until weak assumptions, missing evidence, and useful next steps become visible.
Roundtable is not a fixed prompt pasted into one model. It is a lightweight meeting protocol:
- The Executor reads the task, inspects evidence, and chooses the expert angles needed for this run.
- Lingtai agents take temporary expert roles, surface risks, counterexamples, missing evidence, or explicitly say they have no further objection.
- The Executor resolves disagreement, performs the work, verifies the result, and owns rollback.
This repo ships install scripts, a native Codex skill package, executor-neutral protocol docs, checklists, and examples. Real multi-agent execution still requires Lingtai outside this repo; without Lingtai configured, Roundtable stays in docs/templates mode and must not pretend a real panel ran.
Give this to the coding agent you already use. It should choose the right install path for that terminal:
Install https://github.com/rawpaper123/Roundtable-skill for this project and run the first readiness check. If Lingtai is missing, guide me through installing it. If no Lingtai agent is available, guide me to create or connect at least one. Start with the minimum viable mode: this coding agent as Executor plus one reachable Lingtai agent. If I want the strongest mode, explain which extra provider/API-backed agents to add, but never ask me to paste secrets into chat and never claim a real roundtable ran until the agents reply.
If the check reports docs_only, do not fake expert replies. Configure Lingtai and at least one agent first.
Minimum viable mode is one Executor plus one reachable Lingtai agent. Stronger mode adds more agents from different providers only after the user explicitly configures credentials in the proper local/runtime environment.
There is no universal native "skill" format across all coding agents yet. Roundtable therefore ships a native Codex installer plus an executor-neutral protocol prompt for Claude Code, Cursor, Windsurf, Kimi Work, and other agents. The important requirement is not Codex; it is a working Executor plus at least one reachable Lingtai agent.
When you need manual commands, install paths, or troubleshooting, use the full docs:
Users do not need to hand-pick expert roles. The Executor reads the task, chooses the smallest useful panel, asks for concerns or no-opinion, then owns the final decision. The icons below mark common lead perspectives; the actual panel is assigned per task.
 Development |
Release gates, production bugs, migrations, and risky refactors. |   Reliability, security, data integrity, rollback. |
 Research |
Questions where evidence conflicts or one summary would be too narrow. | Practitioner, scholar, skeptic, incentives, history. |
 Business |
Plans, launches, pricing, partnerships, and go-to-market choices. |   Customer reality, operator cost, finance, legal/risk. |
 Daily |
Personal decisions where reversibility, time, and tradeoffs matter. | Practical friend, budget/time, risk, contrarian. |
More patterns: Use cases, Showcase, Demo script.
- Executor inspects the task and current evidence.
- Executor assigns temporary expert roles to available Lingtai agents.
- Agents reply with must-fix issues, concerns, or
No opinion from my expert perspective. - Executor maps conflicts and decides what advice to accept or reject.
- Executor performs the work or produces the final brief.
- Executor reports evidence, validation, remaining risk, and rollback when relevant.
Use Roundtable first on a real but low-risk task:
- review a small PR before merge,
- pressure-test a research summary,
- review a launch checklist,
- critique a business plan,
- compare two product directions.
Do not use the first run for secrets, production data deletion, irreversible migrations, or high-stakes decisions without human review.
Need help? Use Discussions or open a setup help issue.
Good contributions make Roundtable easier to run in real projects: clearer setup paths, better executor adapters, safer examples, and sharper use-case prompts. Start with CONTRIBUTING.md or open an issue if a first run gets stuck.
MIT