Design, validate, and plan a startup from scratch. Covers market research, competitive analysis, business model, brand identity, product definition, financial projections, and validation experiments. Trigger when the user has a startup idea to explore, wants to validate a business concept, needs a b

Startup Design A structured, multi phase skill that takes a startup idea from raw concept to validated design. It produces a complete set of markdown documents organized by domain, with built in progress tracking so work survives session interruptions. How It Works The process has 8 numbered phases executed sequentially, plus a Pre Flight Check (Phase 0.5) and a Customer Discovery gate (Phase 3.7). Each phase produces output files and updates the progress tracker. If a session is interrupted, resume from the last completed checkpoint. Modes Full Mode (default): Execute all phases in order, including Pre Flight and Customer Discovery. Best for thoroughly designing a startup from scratch. Fast Track Mode: When the user says they want a "quick validation," "rapid assessment," or similar, or when time/budget is clearly limited, run a compressed version: 1. Phase 0.5 (Pre Flight Check) — always run, takes 5 minutes 2. Phase 1 (Intake) — shortened to 1 round of questions; capture any prior customer conversations 3. Phase 2 (Brainstorm) — 3 variations instead of 5 8 4. Phase 3 (Research) — Wave 1 + Wave 2 only (skip customer voice and distribution deep dives) 5. Phase 3.5 (Research Gate) — go/no go checkpoint 6. Phase 3.7 (Customer Discovery) — if founder has 5+ prior conversations, document them; if not, run at least 3 interviews before proceeding 7. Phase 4 (Strategy) — Lean Canvas only 8. Skip Phase 5 (Brand) and Phase 6 (Product) 9. Phase 7 (Financial) — Revenue model only, Stage A (assumption based), no full projections 10. Phase 8 (Validation) — Scorecard + top 3 experiments only Fast Track produces fewer files but still gives the founder a clear go/no go signal with evidence. Note in PROGRESS.md that Fast Track mode was used, so a future session can expand to full mode if the idea passes validation. Language Default output language is English . If the user writes in another language or explicitly requests one, use that language for all outputs instead. Reference: Read references/output guidelines.md once at the start. It defines the standard file header/footer (title, date, phase, confidence, flags), cross phase referencing format, quality examples of good vs. bad output, and how to handle mid process pivots. Phase 0: Resume Check Before anything else, check if a PROGRESS.md file exists in the working directory (or a project subdirectory). If it does, read it and resume from the last incomplete phase. Tell the user: "I found progress from a previous session. You completed [phases]. Picking up from [next phase]." If no progress file exists, start from Phase 1. Phase 0.5: Pre Flight Check Before investing time in the full process, run a fast sanity check — 2 3 targeted searches, 5 minutes maximum. The goal is to surface any immediately disqualifying signals so the founder knows them upfront. Run these three checks: 1. Dominant solution check — Does a well funded, widely adopted solution to this exact problem already exist? Search: "{problem domain} software" , "{problem} tool site:producthunt.com" , "{problem} app reviews" . If a clear market leader with 10k+ customers exists, flag it immediately — this is not a reason to stop, but the founder needs to know the competitive reality before starting. 2. Precedent failure check — Has a company tried this exact idea and failed publicly? Search: "{startup idea} startup failed" , "{problem} startup shutdown" , "why {product category} failed" . Prior failures are not disqualifying — they're learnings. But unknown prior failures are landmines. 3. Regulatory/legal instant kill — Is there an obvious legal reason this idea cannot exist? (E.g., specific financial regulations, data privacy laws in the target geography, licensing requirements.) A quick search prevents building toward a wall. Output: A short message to the founder (3 5 bullet points max) with what was found. Use this format: Keep it brief. This is a heads up, not a full analysis. The project directory and PROGRESS.md don't exist yet at this point — present the findings in the conversation, then save them to {project name}/00 intake/preflight.md during the Phase 1 output step, once the project directory is created. Phase 1: Intake Interview The quality of everything downstream depends on how much context you extract now. Don't rush this — a thorough intake saves hours of misdirection later. Reference: Read references/intake questions.md for the full question set (idea, founders, market, business, constraints), the hard questions that surface blind spots, and interviewing technique. Cover all five question areas plus the hard questions — they set the tone for the entire process and signal that this is an honest assessment, not a cheerleading session. Ask 3 5 questions at a time in a conversational flow, probe vague answers, and after 2 3 rounds summarize what you've understood and ask the user to confirm or correct. Output Save the consolidated intake to {project name}/00 intake/brief.md with all captured information organized clearly. The project name should be derived from the startup idea (kebab case, e.g., pet health tracker ). Create PROGRESS.md at the project root with: project name, start date, language, a checklist of all phases including Pre Flight (0.5) and Customer Discovery (3.7) — mark Pre Flight and Phase 1 complete — and a Notes section for session state. Also save the Pre Flight findings from Phase 0.5 to 00 intake/preflight.md now that the directory exists. Phase 2: Brainstorm Before diving into research, explore the idea space. This prevents premature convergence on the first version of the idea. Process 1. Diverge — Generate 5 8 variations of the core idea. Push boundaries: What if the target market was completely different? What if the business model was inverted? What if you solved a smaller/larger version of the problem? What adjacent problems could you solve instead? What would the "10x version" look like vs. the "simplest possible version"? 2. Analyze — For each variation, note: What's exciting about it What's risky or hard How it changes the competitive landscape 3. Converge — Present the variations to the user. Help them identify which elements resonate. The goal isn't to pick one variation — it's to enrich the original idea with insights from the exploration. 4. Refine — Based on the user's reactions, crystallize the refined idea. Update the brief if the idea evolved significantly. Output Save to {project name}/00 intake/brainstorm.md . Update PROGRESS.md. Phase 2.5: Research Depth Assessment After intake (and brainstorm if applicable), assess market complexity and present the Research Depth recommendation to the user. Reference: Read references/research scaling.md for the complexity scoring matrix, tier definitions, wave configurations, and the user communication template. Process 1. Score three factors from the intake: market breadth (1 3), known competitors (1 3), geographic scope (1 3) 2. Sum the scores (range 3 9) and map to a tier: Light (3 4), Standard (5 7), Deep (8 9) 3. Present the Research Depth table to the user (see research scaling.md for the exact template) 4. Wait for user response: light , deep , or ok to accept the recommendation 5. Record the selected tier in PROGRESS.md The selected tier determines the number of agents per wave and search rounds per agent in Phase 3. See research scaling.md for exact wave configurations per tier. Phase 3: Market Research This is the most resource intensive phase. It uses 4 sequential waves of web research, each building on the previous one's findings. Environment Detection Check if the Agent tool is available (Claude Code) or not (Claude.ai, other environments): Agent tool available: Spawn subagents in parallel within each wave, as described below. This is faster (~3 5 min per wave). Agent tool NOT available (Claude.ai, web): Execute the research yourself, sequentially. For each wave, follow the same agent templates from the reference files, but run the searches one at a time in the main conversation. Cover the same topics and apply the same research principles — the output quality should be identical, it just takes longer. Do NOT skip any wave or reduce search depth because of the sequential mode. Web Search Availability Phase 3 requires WebSearch. In Claude Code, the tool is always available — if the user hasn't pre approved it, the system will prompt them for each search. If the user denies permission, or in environments where WebSearch doesn't exist at all, fall back to Knowledge Based Research Mode : use your training data, clearly mark all findings with [Knowledge Based — not live data, verify independently] , reduce confidence ratings by one level, and recommend the founder verify key claims manually. Note the mode in PROGRESS.md so future sessions know the research wasn't web sourced. References — Read the relevant file for each wave: references/research principles.md — Cross cutting rules (source quality, cross referencing, quantification, handling search failures). Read this FIRST. references/research wave 1 market.md — Agent templates for Wave 1 (market sizing, trends, regulatory) references/research wave 2 competitors.md — Agent templates for Wave 2 (direct, indirect, GTM analysis) references/research wave 3 customers.md — Agent templates for Wave 3 (customer voice, demand, audience) references/research wave 4 distribution.md — Agent templates for Wave 4 (channels, geographic entry) references/research synthesis.md — How to synthesize raw findings into final deliverables Read only the principles file + the wave file you're currently executing. Don't load all wave files at once. Research Principles Each agent performs 5 8 web searches minimum , drilling deeper with each round Cross reference every key finding across 2 3 independent sources Rate source quality (Tier 1: analyst reports, Tier 2: tech press, Tier 3: blogs/social) Quantify everything — "$4.2B at 12.3% CAGR" not "the market is growing" Date all data and flag anything older than 18 months Note contradictions between sources rather than picking one Research Waves Full agent briefs, search strategies, and output destinations live in the wave reference files — read the relevant file when spawning each wave: Wave 1: Market Landscape (3 agents) — A1 Market Sizing & Economics, A2 Industry Trends & Timing, A3 Regulatory & Compliance (skip A3 if no regulatory exposure) Wave 2: Competitive Analysis (3 agents) — B1 Direct Competitor Deep Dives, B2 Indirect Competitors & Substitutes, B3 Competitor Go to Market Wave 3: Customer & Demand (3 agents) — C1 Customer Voice & Pain Points, C2 Demand Signals & Market Validation, C3 Target Audience Profiling Wave 4: Distribution & Partnerships (2 agents) — D1 Distribution Channels, D2 Geographic & Market Entry Each wave must complete before the next starts. Pass key findings forward as context (Wave 1 findings to Wave 2 agents, competitor list and GTM findings to Wave 3, and so on). Agents run in parallel within a wave, or as sequential research blocks without the Agent tool. Raw → Synthesized All agents save raw findings to {project name}/01 discovery/raw/ . After all waves complete, synthesize into 4 polished deliverables. Synthesis is reasoning, not formatting — it's where the raw research becomes a decision. Before writing anything, think hard about how the pieces fit together: which sources conflict and which to trust, what the evidence actually supports versus what the founder hopes, and what it all means for this specific startup. This is the highest leverage thinking in the whole process and every downstream phase inherits its quality, so if the model supports extended thi

startup-design