Research-informed ideation (mandatory)
Execute 5 phases in sequence. Phases 2–5 are pure reasoning (no tool calls).
Phase 1 — DISCOVER: Real-world signal scan
The sole purpose is to gather raw material about real problems people have right now.
- Primary research (web search — this drives ideation):
- Conduct 2–3 targeted web searches to surface:
- Acute pain points and workflow breakdowns across domains (dev tools, productivity, creative work, education, business ops, personal life)
- Recent AI product launches, failures, and community frustrations from the last 7–14 days — what's generating heat, what's disappointing users, what gaps people are vocal about
- Manual or tedious processes ripe for AI-native reinvention — expensive services that could be democratized, expert tasks that could be made accessible
- If web search is unavailable or rate-limited, do not block the run. Fall back to reasoning from training knowledge about real-world pain points and market gaps. Report the blocker in progress output and continue.
- Deduplication fetch (API — not for inspiration):
- Fetch latest 100 approved Sundai projects from
https://www.sundai.club/api/projects?status=APPROVED
- Extract titles and one-line descriptions into a dedup list used only in Phase 3 to filter out overlapping ideas
- Do not analyze these projects for engagement patterns, themes, or inspiration
- Phase 1 output — "Signal Brief":
- 5–10 specific real-world pain points with context (who suffers, what breaks, why current solutions fail)
- 3–5 market shifts or emerging opportunities creating new solution space
- Sundai dedup list (titles + descriptions, for later filtering only)
Phase 2 — DEFINE: Problem space framing
Before generating any solutions, define problems worth solving.
- From the signal brief, select 5 high-potential problem spaces where pain is acute, timing is right, and AI can provide a step-change improvement.
- For each, write a "How Might We" statement: "How might we help [specific person] who [concrete situation/pain] to [desired outcome]?"
- For each, write a Job-to-Be-Done statement: "When [situation], I want to [motivation], so I can [outcome]."
- For each, articulate the current alternative and its gap: "Today they [workaround/existing tool], but it [specific limitation — slow, expensive, requires expertise, error-prone, etc.]."
- Apply a feasibility gate: Can this problem be meaningfully addressed with a demoable AI-powered MVP built in a single autonomous run using free models? Discard any that cannot.
- Phase 2 output: 3–5 validated problem frames, each with HMW + JTBD + current alternative + gap + feasibility confirmation.
Phase 3 — DEVELOP: Divergent idea generation
Generate a broad set of solution concepts using multiple creative lenses, then cull aggressively.
- Divergent pass — generate 6–15 raw concepts:
- For each of the 3–5 problem frames, generate 2–3 solution concepts. Each concept must come from a different creative lens (do not reuse the same lens consecutively):
- Analogical transfer: What solves a structurally similar problem in a completely different domain? Transpose that pattern into an AI-native solution.
- 10x lens: What would a solution look like if it were 10x better than the current alternative, not 10% better? What would have to change fundamentally?
- First principles decomposition: Strip the problem to irreducible atoms. What is the minimum thing that must happen for the user to get value? Build up AI-native from the ground up.
- Constraint-based invention: Force an unusual constraint ("must deliver value in under 10 seconds," "no text input," "combine two unrelated AI capabilities," "works with a single API call") and let it drive novel design.
- Convergent pass — cull to 5 candidates via hard gates:
- Not buildable as a demoable MVP in one autonomous run → cut
- Too similar to an existing Sundai project on the dedup list → cut
- AI is not the core value driver (could work equally well without AI) → cut
- Is a generic chatbot, wrapper, or "ask AI about X" pattern with no distinctive interaction model → cut
- Cannot be understood from a title and one sentence → cut
- Phase 3 output: 5 candidate concepts, each with: working title (≤32 chars), one-line description (≤100 chars), which problem frame it addresses, which creative lens generated it, why AI is essential.
Phase 4 — DELIVER: Scoring and stress-testing
- Pass 1 — Rubric scoring (1–5 each, max 35):
- Real-world urgency: How acute is the underlying pain? Are people actively seeking solutions?
- Novelty: Is this genuinely new — not a rehash of common AI demo patterns (chatbot, summarizer, image generator)?
- Demo wow factor: Would someone say "whoa" within 2 minutes of trying it?
- Feasibility: Can this be built, deployed, and demoed in one autonomous run with free AI models?
- Value clarity: Can someone immediately understand the benefit from the title and a 10-second interaction? Feature → capability → benefit → outcome must be self-evident.
- Job fit: Does this clearly accomplish a job someone is actively trying to do? Would a real person switch from their current approach?
- Shareability: Would someone show this to a friend or post about it? Is there a natural "you have to see this" moment?
- Pass 2 — Pre-mortem on top 2 candidates:
- "Imagine nobody tries the demo and it gets zero engagement. What went wrong?" — 3 specific failure reasons each.
- "What is the single riskiest assumption this idea depends on?" — one sentence.
- "Can the MVP design itself mitigate that risk?" — yes/no with a concrete mechanism.
- Selection: Pick the candidate that scores highest AND survives pre-mortem without a fatal, un-mitigable flaw. If top scorer has a fatal flaw, take the runner-up. If both finalists have fatal flaws, return to Phase 3 pool.
- Selected idea must meet all of: clearly differentiated from common AI demo patterns; demoable in minutes by a first-time user; creative and ambitious; rooted in a real pain point surfaced during research; AI is essential and central.
- Override: If user already gave a fixed idea, skip Phases 1–4 and proceed directly to Phase 5.
Phase 5 — PACKAGE: Structured handoff to Step 2
Format the selected idea for direct consumption by Step 2 when writing spec/spec.md:
IDEATION OUTPUT
Title: [≤32 chars]
Description: [≤100 chars]
Spec Seed:
- What it does: [2–3 sentences]
- Who it's for: [specific target user and their situation]
- Core job: [JTBD one-liner from Phase 2]
- Current alternative: [what people do today and why it falls short]
- Key differentiator: [the single thing that makes this unlike anything else]
- AI integration: [how AI is core to the product's value — which capability, how surfaced to user]
- Demo flow: [3–5 step user journey for the smoke test in Step 11]
- Tech stack suggestion: [frontend + backend + AI provider/model]
- Riskiest assumption: [from pre-mortem — what could kill this + how MVP mitigates it]
Engagement hook: [one sentence — why someone would click and try the demo]
Design Direction:
- Visual style: [one of: minimal, bold, playful, editorial, dashboard]
- Reference design system: [pick from references/design-systems.md — e.g., "Dev Tool (Linear-inspired)" for dev tools, "Fintech (Stripe-inspired)" for business tools, "AI Platform (Claude-inspired)" for AI apps]
- Color palette: [pick from references/design-palettes.md or specify custom — list 5 hex colors: primary, secondary, accent, background, text]
- Font pairing: [heading font + body font from Google Fonts]
- Layout: [single page scroll / sidebar dashboard / multi-page app]
- Key polish target: [the one screen or interaction that must feel premium]
- Title ≤32 chars and description ≤100 chars are hard limits enforced by Sundai.
- Design Direction must reference a specific design system and palette — do not leave generic or "TBD". The Fabro build and polish steps use these values to produce polished UI.