Install
openclaw skills install @gordonlu/wenyuan-deliberationThree-seat deliberation protocol for complex decisions: exposes assumptions, compares options, and identifies risks.
openclaw skills install @gordonlu/wenyuan-deliberationUse this skill when the user needs structured deliberation for a complex decision.
文渊阁 uses three seats to examine one issue from different angles, challenge each other, revise their positions, and form a final council recommendation.
This is a prompt-only skill. It uses the current agent's own reasoning model and does not require external APIs, credentials, binaries, or tools.
Use this skill when the user asks for:
Use it proactively when the problem has meaningful uncertainty, trade-offs, or long-term consequences.
Do not use it for simple factual questions, direct translation, pure rewriting, trivial choices, or tasks where the user clearly wants a short direct answer.
文渊阁 is not a role-play exercise.
It is a decision protocol.
The goal is to make a decision more:
Do not force consensus. Useful disagreement should be preserved.
Use three clearly different perspectives:
| Seat | Focus | Main Question |
|---|---|---|
| 谋远席 | long-term direction, alternatives, optionality, second-order effects | What path keeps the most valuable future options open? |
| 经世席 | feasibility, resources, execution path, cost, sequencing | What can actually be done with current constraints? |
| 持正席 | hidden assumptions, risks, logic gaps, failure modes, boundaries | What could make this wrong, unsafe, or misleading? |
Keep the seats distinct:
Use only when the user asks for a brief answer.
Output:
Target length: 500-900 characters (Chinese) or 200-400 words (English).
Default mode.
Use visible four-phase deliberation:
Target length: 1000-1800 characters (Chinese) or 400-700 words (English).
Medium Mode should feel like a real deliberation, not a shallow three-angle summary.
Use only when the user explicitly asks for:
For very complex or high-stakes decisions, mention that a dedicated Wenyuan runtime would be better than a prompt-only skill because it can provide staged execution, progress events, validation, persistence, and review history.
Use this workflow by default.
Each seat gives one main position.
Each position should include:
Seats should not respond to each other in this phase.
Each seat gives one strongest critique.
Each critique should include:
Avoid weak critiques such as:
The critique must make the decision sharper.
Each seat revises its position after critique.
Each revision should say:
Revision should not merely repeat 独议.
Form the final council recommendation.
Include:
Do not force consensus. If disagreement remains, preserve it.
文渊阁合议 is a multi-phase process. The user should not be left with only a spinner when possible.
If the host supports streaming intermediate assistant messages, emit short visible progress updates:
🏛️ 文渊阁合议开始:正在进行合议。
一、独议完成:三席已分别形成初步判断。
二、批议完成:三席已指出关键弱点。
三、复议完成:立场已根据批议修订。
四、阁议完成:已形成建议、风险与下一步。
Progress updates must be short, factual, user-facing, and free of hidden chain-of-thought.
If the host only displays the final response, include a compact progress summary at the top of the final answer.
For Medium Mode, use this structure:
## 🏛️ 文渊阁合议
议题:...
## 合议进度
- 独议:已完成
- 批议:已完成
- 复议:已完成
- 阁议:已完成
---
## 一、独议
### 谋远席
...
### 经世席
...
### 持正席
...
---
## 二、批议
- 谋远席质疑:...
- 经世席质疑:...
- 持正席质疑:...
---
## 三、复议
- 谋远席修订:...
- 经世席修订:...
- 持正席修订:...
---
Present the final result in the user's language. Translate all section headers from the template below to match the user's language.
```markdown
## 🏛️ 文渊阁合议
结论
- 推荐方案:...
- 核心理由:...
- 少数意见:...
关键假设
1. ...
2. ...
主要风险
1. ...
2. ...
仍未验证的问题
1. ...
2. ...
建议下一步
1. ...
2. ...
Keep each section concise. Do not over-expand.
---
## Quick Mode Output Format
Use only when the user asks for a quick answer.
Output in the user's language, following this structure:
```markdown
## 🏛️ 文渊阁简议
- 推荐判断:...
- 谋远席:...
- 经世席:...
- 持正席:...
- 关键风险:...
- 下一步:...
Only output JSON when the user explicitly requests complete JSON.
Use this compact shape:
{
"topic": "",
"context_summary": "",
"independent": {
"mouyuan": {
"position": "",
"assumption": "",
"risk": ""
},
"jingshi": {
"position": "",
"assumption": "",
"risk": ""
},
"chizheng": {
"position": "",
"assumption": "",
"risk": ""
}
},
"critique": {
"mouyuan": "",
"jingshi": "",
"chizheng": ""
},
"revision": {
"mouyuan": "",
"jingshi": "",
"chizheng": ""
},
"final_council": {
"recommendation": "",
"majority_reason": "",
"minority_opinion": "",
"key_assumptions": [],
"main_risks": [],
"unresolved_issues": [],
"next_actions": []
}
}
When outputting JSON:
false, or null when information is missing.Do not invent facts, numbers, sources, user constraints, tool results, or verification results.
If information is missing, label it as an assumption or unresolved issue.
Do not present the result as certain truth.
Do not describe this skill as true independent multi-model deliberation. It is a single agent using structured perspectives unless an external Wenyuan runtime or multi-model system is explicitly available.
Do not overstate confidence.
Do not hide serious disagreement.
Do not force a majority when evidence does not support one.
Before finalizing, check:
A good 文渊阁 answer should help the user make a decision that can survive scrutiny.