Research Ideation

Research Ideation

Overview

Generate concrete research ideas from loose inputs — keywords, datasets, papers — and present them with actionable metadata. This skill is divergent: it expands possibilities. Its downstream partner research-brainstorming is convergent: it narrows one idea into a rigorous study design.

Core principle: You cannot refine a question you haven't asked yet. Ideation comes before design.

When to Use

Use this skill when:

• The user has no formed research question — only keywords, a dataset, or vague interest
• The user says "what should I research?", "what can I do with this data?", "research ideas for [topic]"
• A dataset or paper is provided without a specific question attached
• whats-next diagnoses the state as Pre-ideation

Do NOT use this skill when:

• The user already has a research question (e.g., "Does X cause Y?") → use research-brainstorming
• The user has a hypothesis to test → use hypothesis-first
• The user is stuck mid-project → use whats-next

Discriminating rule: If the user can state their interest as "Does/Is/Can [X] [verb] [Y]?", they have a formed question → skip to research-brainstorming. If they cannot, they are in ideation territory.

Checklist (Step Order Fixed)

You MUST create a task for each of these items and complete them in order:

1. Collect inputs — gather keywords/datasets/papers from conversation or by asking
2. Explore sources — analyze what's available
3. Generate ideas — 3-5 by default; 5-10 if scope is broad
4. Present ideas — each with title + description + metadata table
5. Recommend and propose — highlight one + suggest research-brainstorming handoff

Step 1: Collect Inputs

The skill accepts three types of input. All optional, but at least one must be present:

1. Keywords / interest area — e.g., "attention mechanism", "neuroimaging", "time-series forecasting"
2. Dataset — files in the project (csv, npy, etc.) whose structure the agent explores
3. Papers / literature — user-provided PDFs, URLs, or web search for recent trends

Collection rules:

• Check what the user already provided in conversation. Skip what's already known.
• If nothing is provided, ask once: "Are there datasets or papers I should look at? Or just throw me some keywords."
• Minimum requirement: one keyword. The skill can start from a single keyword alone.
• Do NOT ask multiple questions. One prompt, then start working.

Step 2: Explore Sources

Adapt exploration to the input types available:

| Input type | What to do | |------------|------------| | Dataset | Read the file(s). Analyze columns, shape, dtypes, descriptive stats (N, mean, std, NaN rate, label distribution). Note what variables are available and what relationships could be studied. | | Paper | Read abstract + methodology + limitations. Extract the research question, key findings, and identified gaps. Note what the authors suggest for future work. | | Keywords | Web-search for recent trends, open problems, and active debates in the area. Look for survey papers, workshop topics, and recent preprints. |

When multiple input types are present, look for cross-pollination — where dataset characteristics meet literature gaps or keyword trends.

Step 3: Generate Ideas

Default: 3-5 ideas. If the agent judges the scope is broad (multiple fields, large dataset with many variables, or diverse literature), expand to 5-10 ideas without asking.

Each idea should be:

• Concrete — specific enough that research-brainstorming could start refining it
• Distinct — ideas should not be minor variations of each other
• Grounded — connected to the actual inputs (data available, literature gap identified, trend observed)

Step 4: Present Ideas

Each idea follows this template:

N. [Idea Title]

[2-3 sentences: why this is interesting, what gap it fills]

| Item | Detail | |------|--------| | Difficulty | High / Medium / Low | | Data needed | Already available / Additional collection required / Public dataset available | | Estimated duration | ~2 weeks / ~1 month / ~3 months (one researcher, full-time, excluding data collection) | | Core methodology | e.g., "transformer + EEG temporal features" |

Step 5: Recommend and Propose

After presenting all ideas, recommend one:

> "I find #N most promising — [one-sentence reason]. Want to start research-brainstorming to shape this into a rigorous study design?"

The user may:

• Pick the recommended idea → invoke research-brainstorming
• Pick a different idea → invoke research-brainstorming with that idea
• Ask for more ideas → generate additional ideas
• End the session → suggest journaling (see Session End below)

All are valid outcomes. The handoff is a suggestion, never a forced transition.

Session End Behavior

If the user ends the session without picking an idea:

> "Want me to save these ideas before we wrap up? research-journal can capture them so you don't lose them."

This is a gentle suggestion. Ideas are ephemeral by default — saved only if the user journals or picks one to develop.

Anti-Patterns

| What you might do wrong | What to do instead | |---|---| | Start designing a study for one of the ideas | Stop. That's research-brainstorming's job. Present ideas, don't refine them. | | Generate only one idea | Always generate at least 3. The user needs options. | | Ask 5 questions before generating | Collect inputs in one prompt, then work. | | Force the user to pick an idea | Present, recommend, wait. The user decides. | | Generate ideas unrelated to the inputs | Every idea must connect to at least one input (keyword, dataset feature, or literature gap). | | Skip exploring the dataset | If a dataset is provided, you MUST look at its structure. Don't ideate in the abstract. |

Integration

• Called by: eureka:using-eureka (when no formed question exists), eureka:whats-next (Pre-ideation diagnosis)
• Hands off to: eureka:research-brainstorming (when user selects an idea — suggestion only)
• Does NOT invoke: hypothesis-first, experiment-design, or any execution/review skill

Skill Type

FLEXIBLE — The step order (collect → explore → generate → present → recommend) is fixed and must not be skipped. But how each step executes adapts to context:

• Dataset-heavy input → more structural analysis, fewer web searches
• Keyword-only input → heavier web search for trends and gaps
• Broad field → 5-10 ideas; narrow niche → 3-5 ideas

No scientific integrity is at stake at the ideation phase. The downstream rigid skills (hypothesis-first, claims-audit) enforce that discipline later.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jeonnoin-alt
• Source: jeonnoin-alt/Eureka
• License: MIT
• Homepage: https://github.com/jeonnoin-alt/Eureka

Install and usage instructions live in the source repository linked above.