| name | idea-generator |
| description | Use when the user wants to generate or refine one research idea from a required Markdown brief and local Markdown paper notes stored in `papers/**/*.md`, following `prompts/generate_idea.md`. |
| version | 0.2.0 |
Idea Generator
Generate exactly one stronger and more defensible research idea by combining:
- a required user-authored Markdown brief provided through
input_md_path
- local paper notes stored as Markdown files under
papers/**/*.md
- the authoritative generation template in
prompts/generate_idea.md
When to Use
Use this skill when:
- the user already has a research direction, problem statement, or constraint set written in Markdown
- the repository contains local paper notes in
papers/**/*.md
- the goal is to propose one better idea grounded in those paper methods instead of doing open-ended brainstorming
Required Inputs
input_md_path (required)
The caller must provide a workspace-relative or absolute path to a user-authored .md file.
The file should contain as many of the following as possible:
- research direction or problem statement
- target scenario or task setting
- hard constraints and non-goals
- observed pain points or hypotheses
- optional target venue, dataset, metric, or implementation preference
papers/**/*.md (required local paper source)
Use local Markdown paper notes as the only paper source for this skill.
Rules:
- read only the relevant subset
- do not invent missing details
- if the corpus is weak or empty, say so explicitly
prompts/generate_idea.md (authoritative template)
Always read prompts/generate_idea.md before producing the final answer.
Treat it as the authoritative prompt for:
- reasoning order
- output structure
- idea quality bar
Workflow
-
Validate input_md_path.
- Confirm the file exists and is Markdown.
- If it is missing or unreadable, stop and ask the user for a valid path.
-
Read the user brief first.
- Extract the problem, scenario, constraints, assumptions, non-goals, and what “better” should mean.
-
Scan papers/**/*.md.
- Find the smallest relevant subset.
- Extract method family, mechanism, assumptions, strengths, limitations, and reusable design patterns.
-
Identify the improvement opportunity.
- Compare the brief against the local paper map.
- Pin down what should be reused, what should change, and what contradiction or failure mode the new idea should resolve.
-
Use paper-miner only when helpful.
- Invoke it if the local notes are numerous, heterogeneous, or weakly structured.
- Otherwise complete the task locally.
-
Load prompts/generate_idea.md.
- Follow its workflow and output structure exactly.
-
Produce exactly one coherent idea.
- Prefer conceptual unity over module stacking.
- Prefer the minimal number of tightly coupled contributions.
- Make clear what is reused, what is changed, and why the result is more reasonable than the source methods.
Source Rules
- Base every core judgment on
input_md_path and relevant papers/**/*.md.
- Treat local paper notes as grounded evidence for ideation.
- Do not claim a paper supports something unless that support appears in the local note.
- Do not fall back to online search inside this skill.
Output Rules
- Respond in Chinese.
- Follow the structure required by
prompts/generate_idea.md.
- Keep assumptions and uncertainty inside the required sections.
- State integration scope, likely compute or data cost, and key assumptions explicitly.
Failure Handling
- Missing
input_md_path: ask the user for a valid Markdown path.
- Empty or irrelevant
papers/**/*.md: say that local paper grounding is unavailable, then stop or proceed with reduced confidence.
- Conflicting paper notes: surface the conflict before committing to the idea.
- Weak novelty evidence: avoid strong claims such as
first, novel, or state of the art.