Research Idea Convergence

Research Idea Convergence

Canonical Summary

Interactive research direction selection for academic projects. Generates multiple candidate research directions from survey/literature results, presents them with structured comparison for user selection, then converges to a final publishable angle.

Trigger Rules

Use this skill when:

• The project enters the ideation stage after completing the survey stage
• The user asks about research direction, publishable angle, or idea selection
• The current task involves generating or selecting a research direction

Do NOT use this skill when:

• The user has already decided on a specific research direction and just wants to refine it
• The project is still in the survey stage (use survey skills instead)

Execution Contract

• Resolve every relative path from this skill directory first.
• Do not write generated artifacts back into the skill directory; save them inside the active project workspace.
• Never skip the user selection checkpoint — this is the core purpose of this skill.

Working Rules

1. Read existing survey outputs before generating candidates.
2. Generate 2-4 candidate directions, no more, no fewer.
3. Each candidate must be concrete enough to evaluate (not vague platitudes).
4. [USER_CHECKPOINT] — Always stop and wait for user selection. Never auto-select.
5. Only after user confirmation, write the final direction to output files.

Step-by-step Instructions

Step 1 — Read Survey Context

Read all available survey outputs to understand the research landscape:

File	Purpose
.pipeline/docs/research_brief.json	Research topic, goal, target venue
.viewerleaf/research/Survey/reports/*.md	Literature screening, gap summary
.pipeline/tasks/tasks.json	Current task states and progress

Extract:

• Research domain and specific problem area
• Key gaps identified in the literature
• Strong baselines that exist (traditional and learning-based)
• Available datasets / experimental platforms
• Target venue constraints (if specified)

Step 2 — Generate Candidate Directions

Produce 2-4 candidate research directions. For each candidate, provide:

## 候选方向 N：[方向名称]

**核心思路：** [1-2 句技术路线和创新点]

**研究问题：** [具体要回答的科学问题]

**方法概要：**
- [关键技术 1]
- [关键技术 2]
- ...

**对标基线：**
- [需要对比的已有方法 1]
- [需要对比的已有方法 2]

**目标刊物：** [适合投稿的会议/期刊]

**✅ 优势：**
- [优势 1]
- [优势 2]

**⚠️ 风险：**
- [风险 1]
- [风险 2]

**可行性评分：** ⭐⭐⭐⭐☆ (x/5)
- 实验资源: x/5
- 时间可控: x/5
- 技术难度: x/5
- 发表潜力: x/5

Candidate Generation Principles

1. Differentiation — Candidate directions should be clearly distinct (different technical routes, different innovation angles, different application scenarios)
2. Comparability — Use unified evaluation dimensions to facilitate horizontal comparison by the user
3. Pragmatism — Consider the user's actual resources (undergraduate vs. master's/doctoral, experiment cycle, computational resources)
4. Risk Gradient — Include both conservative and safe directions and aggressive and innovative directions, letting the user choose between risk and reward

Step 3 — Present Comparison Table

After listing all candidate directions, provide a summary comparison table:

## 📊 候选方向对比

| 维度 | 方向 1 | 方向 2 | 方向 3 |
|------|--------|--------|--------|
| 创新性 | ⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ |
| 可行性 | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐ |
| 发表潜力 | ⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ |
| 实验周期 | 2 周 | 4 周 | 6 周 |
| 推荐度 | 🟡 稳妥 | 🟢 推荐 | 🔴 激进 |

Step 4 — USER CHECKPOINT ⛔

[USER_CHECKPOINT] — You must stop here and wait for the user's reply.

Ask the user:

1. Which direction do you lean toward?
2. Do you need to combine elements from certain directions?
3. Do you have any of your own ideas you want to add?
4. Do you have any concerns about the risks of a certain direction?

Absolutely do not skip this step. Do not make the choice for the user.

After the user makes a clear choice, proceed to Step 5.

Step 5 — Converge and Write

After the user confirms:

1. Write to publishable_angle.md

   路径: .viewerleaf/research/Ideation/publishable_angle.md
   

   Include: the complete description of the selected direction, the user's revision comments (if any), the benchmark baseline, and the expected contribution

2. Update research_brief.json

   路径: .pipeline/docs/research_brief.json
   

   Update briefGoal and stageNotes.ideation

3. Update tasks.json

   路径: .pipeline/tasks/tasks.json
   

   Mark the ideation task as complete and trigger the next-stage task

4. Record the rationale for the choice At the end of publishable_angle.md , append:

   ## 选择记录
   - 候选方向数：N 个
   - 用户选择：方向 X [+ 用户修改内容]
   - 选择理由：[用户给出的理由或综合分析]
   - 弃选方向：方向 Y (原因)、方向 Z (原因)
   

Output Templates

publishable_angle.md

Use templates/publishable_angle.template.md:

# Publishable Angle

## 研究方向
[选定方向名称]

## 核心创新点
[1-2 句话概括]

## 研究问题
[具体要回答的科学问题]

## 方法概要
- ...

## 对标基线
- ...

## 预期贡献
- ...

## 目标刊物
[期刊/会议名称]

## 选择记录
- 候选方向数：N 个
- 用户选择：方向 X
- 选择理由：...

Integration with Other Skills

Upstream	This Skill	Downstream
research-literature-trace (survey reports)	→ Read gap/baseline information
research-pipeline-planner (brief)	→ Read the research topic and goals
	Generate candidates → user selection → write	→ inno-idea-eval (review the selected direction)
		→ inno-experiment-dev (experimental design)

Notes for LLMs

1. Core Principles: The value of this skill lies ininteraction, not in generation. Generating candidates is only a means; enabling the user to make an informed choice is the goal.
2. Do not skip USER_CHECKPOINT: Even if the user appears to already have a preference, fully present all candidates and explicitly collect confirmation.
3. Do not over-recommend: You may label something as "recommended," but do not imply that there is only one correct answer. Each direction has its applicable scenarios.
4. Respect the user's existing ideas: If the user already has preliminary ideas, include them as one of the candidates in the comparison rather than ignoring them.
5. Differentiation is very important: Do not generate several directions that differ only slightly; that is equivalent to giving no choice. Technical routes, innovation angles, and target scenarios should be clearly distinguished.