| name | vs-research-ideate |
| description | Generate high-entropy research (자료조사) and ideas (아이디어) using Verbalized Sampling to avoid mode collapse and maximize creativity and novelty. |
This skill guides the creation of distinctive, non-generic research (자료조사) and ideas (아이디어) by explicitly mitigating "Mode Collapse"—the tendency to produce obvious, first-page-of-Google results and predictable "AI-slop" ideas. It uses Verbalized Sampling (VS) logic to unlock LLM creativity and deliver high-entropy, memorable research angles and ideation.
Use when: The user asks for research, literature review, market/competitor analysis, background gathering, or ideation, brainstorming, concept generation, feature ideas, or creative solutions.
Phase 0: Context Discovery (MANDATORY)
BEFORE any research or ideation, you MUST gather deep context from the user. Use the AskUserQuestion tool (or direct questions in dialogue) to probe the following dimensions:
Discovery Dimensions
- Purpose: Why is this research or ideation needed? (e.g., decision-making, report, inspiration, validation, problem-solving)
- Scope & Constraints: What's in bounds? Time, budget, geography, domain? What must be excluded?
- Existing Knowledge: What does the user already know or assume? What have they already read or tried?
- Reference Points / Anti-References: What angles or sources do they want to emulate? What do they explicitly want to avoid? (e.g., "not the usual McKinsey-style framing")
- Audience: Who will consume this research or these ideas? What's their level of expertise and what would surprise or help them?
Context Signals
- Analyze the prompt for implicit scope (industry, product, academic vs commercial, depth of evidence required).
- Infer anti-patterns from domain (e.g., tech: avoid "top 10 tools" generic lists; strategy: avoid generic SWOT).
- Ask follow-up questions if the request is vague—surface the real question or problem behind the ask.
CRITICAL: Do not proceed to Phase 1 until you have sufficient context. A vague prompt requires MORE questions, not assumptions.
Phase 1: Identify the Mode (The Generic Baseline)
- For research: Verbalize the most predictable, high-probability (P ≈ 0.95) research approach and sources for this request. Examples of "research mode collapse":
- First-page Google results and Wikipedia
- Generic industry reports and "state of X" articles
- Obvious keywords and mainstream frameworks (e.g., Porter's Five Forces without twist)
- Single-domain, single-perspective synthesis
- For ideation: Verbalize the most predictable ideas—the first three an average person (or typical AI) would suggest. Examples:
- "Add a dashboard," "improve UX," "use AI"
- Incremental tweaks with no reframe
- Ideas that match every other competitor's playbook
- CRITICAL: You are forbidden from delivering only this baseline. It is the floor to diverge from.
Phase 2: Sample the Long-Tail (Probability Mapping)
Generate three distinct directions (research angles or idea clusters) and assign a "Typicality Score" (T-Score) from 0 to 1.0 (where 1.0 is most generic):
- Direction A (T ≈ 0.7): Solid and credible but safe; mainstream sources or obvious-but-good ideas.
- Direction B (T ≈ 0.4): Distinctive; niche sources, cross-domain parallels, or ideas with a clear point of view.
- Direction C (T < 0.2): Experimental; contrarian evidence, underused primary sources, or bold/reframed ideas that are non-obvious.
T-Score Justification Required: For each direction, explicitly state WHY it has that T-Score. What makes it more or less typical? Reference specific choices (sources, framings, or idea attributes).
Phase 3: Commit to Low-Typicality
Select the direction with the lowest T-Score that still meets:
- Purpose and scope from Phase 0
- All Quality Guardrails (see below)
Commit to this BOLD research path or idea set with intentionality. The choice must be DELIBERATE, not accidental.
Quality Guardrails (NON-NEGOTIABLE)
These principles MUST be satisfied regardless of how experimental the research or ideas become. If a Low-T direction violates any of these, increase T until compliance is achieved.
For Research (자료조사)
| Guardrail | Description |
|---|
| Traceability | Sources and claims can be verified; distinguish fact vs inference vs opinion |
| Relevance | Every piece of evidence or source clearly ties to the user's question or scope |
| Coherence | One clear narrative or argument; no random fact-dumping |
| Actionability | Takeaways are explicit: so what? what should the user do or believe? |
For Ideation (아이디어)
| Guardrail | Description |
|---|
| Clarity | Each idea is concretely described—enough to judge and to take a next step |
| Feasibility | Ideas are scoped; not pure fantasy unless the user asked for blue-sky only |
| Internal Consistency | The set of ideas fits a coherent strategy or theme (e.g., same product, same user) |
| Differentiation | Ideas are distinguishable from each other and from the "mode" (Phase 1) |
Anti-Patterns (EXPLICIT FAILURES)
If your output exhibits these patterns, you have FAILED the skill's intent:
1. Accidental Research / Ideas
- Adding "different" sources or ideas without intentionality
- Randomness masquerading as creativity
- Unable to articulate WHY a source or idea was chosen
- Test: If asked "why this source or idea?", you must have a coherent answer
2. Frankenstein Output
- Mixing incompatible framings or domains without synthesis
- No unifying thread or narrative (research) or strategy (ideas)
- Reads like unrelated fragments from different projects
- Test: Could you describe this research or idea set's "point" in one sentence?
3. Mode Collapse in Disguise
- Claiming "long-tail" while still citing only mainstream sources or offering generic ideas
- T-Scores that are not justified or are clearly understated
- Test: Would an expert in the domain find this obvious or surprising?
Research Guidelines (VS-Enhanced)
Apply the Inversion Principle: If a source or angle feels "obvious," it has too much probability mass. Consider the lower-probability, higher-impact alternative—as long as it stays relevant and traceable.
Source Diversity
- High-T examples: Wikipedia, first-page SERP, generic "best practices" blogs, single-industry reports only
- Low-T approach: Primary sources, niche journals, cross-industry or historical parallels, expert interviews or transcripts, non-English or regional sources when relevant, contrarian or critical takes
Angle and Framing
- Avoid the default framing (e.g., "benefits of X"). Try: downsides, edge cases, who disagrees, what changed in the last N years, what’s missing from the mainstream story.
- Use a clear lens: e.g., "through the lens of risk" or "focusing on adoption barriers" so the research has a point of view.
Synthesis
- Synthesize explicitly: themes, tensions, gaps. Don’t just list sources—say what they add up to and what the user should do with it.
Ideation Guidelines (VS-Enhanced)
- Inversion: If the obvious idea is "add feature X," consider "remove X," "replace X with Y," or "make X optional for a segment."
- Combination: Combine constraints or domains (e.g., "what if we applied X from industry A to our product in industry B?").
- Constraint removal: Temporarily ignore one key constraint (cost, tech, time) to unlock ideas, then map back to feasibility.
- Anti-benchmark: Explicitly avoid ideas that "everyone else" is already doing; label and skip the mode, then push further.
Structural Frameworks (Context-Dependent)
For Research Output
When the goal is decision or recommendation, structure the deliverable as:
| Stage | Goal | Application |
|---|
| Question | State the exact question or decision being informed | One sentence; no ambiguity |
| Evidence | Present findings by theme or source type, with T-Score awareness | Prefer low-T sources; mark confidence and gaps |
| Gaps & Caveats | What’s missing, conflicting, or uncertain | Builds trust and clarifies "so what" |
| Implications & Options | What the user should do or believe; 2–3 concrete options if relevant | Actionable next steps |
For Ideation Output
When the goal is concept or feature generation, structure as:
| Stage | Goal | Application |
|---|
| Problem / Opportunity | Reframe the brief in one line | Ensures ideas solve the right problem |
| Idea Set | 3–5 distinct ideas with T-Score and brief rationale | Lowest T that still passes guardrails |
| Selection Criteria | How to choose (e.g., impact vs effort, strategic fit) | Makes the set usable for the user |
| Next Steps | One concrete next step per idea (e.g., "validate with one customer") | Moves from idea to action |
Implementation Standards
- Output quality: Research must be citable and synthesised; ideas must be specific and scoped.
- Complexity–Typicality balance: As you go lower-T, add clarity and structure so the output remains usable.
- No refusal on novelty: Do not simplify or fall back to generic output for brevity. The user invoked this skill for high-entropy research and ideas.
Execution Process Summary
[Phase 0] Context Discovery
↓ (questions – gather purpose, scope, anti-refs, audience)
[Phase 1] Identify the Mode
↓ (verbalize the generic baseline – research or ideas)
[Phase 2] Sample the Long-Tail
↓ (three directions with justified T-Scores)
[Phase 3] Commit to Low-Typicality
↓ (select lowest T that passes Guardrails)
[Phase 4] Execute Output
↓ (research: sources + synthesis + implications; ideas: set + criteria + next steps)
[Phase 5] Surprise Check
↓ (would an expert find this obvious? if yes, refine)
Final Validation
Before delivering, ask yourself:
- Intentionality: Can I justify every major source or idea choice?
- Coherence: Does the research have one clear narrative? Do the ideas form a coherent set?
- Guardrails: Are traceability, relevance, clarity, and feasibility preserved?
- Surprise: Would this stand out compared to generic research or generic ideation?
REMEMBER: The goal is to maximize "Surprise Score" and usefulness while maintaining rigor and clarity. Break the mold—deliberately—in 자료조사 and 아이디어.