| name | idea-scout |
| description | Deep, autonomous domain decomposition and landscape research for any idea or problem statement. Takes a raw idea (from one sentence to a paragraph) and runs a 6-phase research process: decomposes the idea into problem space, target users, domain, adjacent domains, and assumed solution direction; performs exhaustive parallel web research on current state of the art, key players, recent momentum, technical foundations, and user sentiment; synthesizes gaps, blind spots, friction points, challenges, and opportunities; maps 3-5 distinct solution directions with feasibility and differentiation analysis; runs contrarian and timing checks including cross-domain analogies; and scores all findings by confidence level with full source attribution. Outputs a structured research dossier to /docs/idea-scout/. Use this skill whenever the user has a raw idea they want explored before building, wants to understand a problem landscape, needs domain research, or wants to validate whether an idea is worth pursuing. Triggers on: 'scout this idea', 'research this idea', 'explore this concept', 'what's the landscape for', 'idea-scout', 'is this worth building', 'what exists for', 'explore the space of', 'I have an idea for', 'help me understand the landscape', 'what's out there for', 'domain research for', 'validate this idea'. |
| argument-hint | <idea or problem statement> |
Idea Scout — Deep Domain Decomposition & Landscape Research
Why This Skill Exists
Starting a project with incomplete understanding of the landscape is the most expensive mistake in product development. Teams build things that already exist, miss adjacent innovations they could leverage, and overlook fundamental challenges that kill projects months in. This skill exists to prevent that — it forces a thorough, structured exploration of any idea before a single line of code or PRD is written.
The output is not a plan. It's intelligence — a structured understanding of the terrain so that whatever comes next (a PRD, a prototype, a pivot) is grounded in reality rather than assumptions.
Core Philosophy: Three Layers of Knowledge
Every finding in this dossier belongs to one of three layers (inspired by gstack's Search Before Building ethos):
- Layer 1 — Tried and true. Standard patterns, battle-tested approaches. Valuable because you shouldn't reinvent what works. But question assumptions — sometimes the "obvious" answer is wrong.
- Layer 2 — New and popular. Current trends, recent blog posts, ecosystem momentum. Search for these, but scrutinize — hype is not validation. The crowd can be wrong about new things.
- Layer 3 — First principles. Original insights derived from reasoning about THIS specific problem. These are the most valuable findings of all. When first-principles reasoning contradicts conventional wisdom, that's a EUREKA moment — name it, celebrate it, build on it.
The best outcome of research is not finding a solution to copy. It's understanding what everyone does and WHY (Layers 1+2), then applying first-principles reasoning to their assumptions (Layer 3) to discover why the conventional approach is wrong HERE.
How It Works: Eight Phases, Fully Autonomous
The process runs without pausing for user input. Each phase feeds the next. Parallel agents handle research-intensive phases.
User's Raw Idea
|
v
+--- PHASE 1: DECOMPOSE ----------------------+
| Break idea into 5 dimensions + extract |
| falsifiable premises |
+----------------------------------------------+
|
v
+--- PHASE 2: LANDSCAPE RESEARCH -------------+
| 5 parallel agents + api-database-scout |
| Structured by 3 knowledge layers |
+----------------------------------------------+
|
v
+--- PHASE 3: STATUS QUO DEEP DIVE -----------+
| Map the real competitor: current workarounds |
| people cobble together today |
+----------------------------------------------+
|
v
+--- PHASE 4: GAP & OPPORTUNITY ANALYSIS -----+
| Blind spots, Friction, Challenges, |
| Opportunities, Demand reality check |
+----------------------------------------------+
|
v
+--- PHASE 5: SOLUTION SPACE MAPPING ---------+
| 3-5 directions + narrowest wedge |
| + combinatorial opportunities |
+----------------------------------------------+
|
v
+--- PHASE 6: CONTRARIAN & TIMING CHECK ------+
| Cross-domain analogies, bear case, |
| timing analysis, eureka moments |
+----------------------------------------------+
|
v
+--- PHASE 7: DOSSIER REVIEW -----------------+
| Independent subagent reviews for |
| completeness, consistency, gaps |
+----------------------------------------------+
|
v
+--- PHASE 8: COMPILE & VERDICT --------------+
| Structured dossier + opinionated verdict |
| BUILD / DON'T BUILD / PIVOT |
+----------------------------------------------+
|
v
Research Dossier in /docs/idea-scout/
Execution Instructions
Architecture: File-Based Coordination
Like api-database-scout, this skill uses file-based coordination to stay within token limits. Research agents write findings to /tmp/idea-scout/{idea-slug}/. You (the orchestrator) track progress and delegate — you never accumulate raw research in your own context. A final compiler reads all files and produces the dossier.
Step 0: Capture and Acknowledge
Take the user's idea (available as $ARGUMENTS or from the conversation) and acknowledge it:
Scouting: "{user's idea}"
Running 6-phase deep research. This will take several minutes.
I'll report back with a complete research dossier.
Derive a short kebab-case slug from the idea (e.g., "plant health tracking app" -> "plant-health"). Create the working directory:
mkdir -p /tmp/idea-scout/{idea-slug}
Also prepare the output directory:
mkdir -p docs/idea-scout
Phase 1: DECOMPOSE — Break the Idea Into Dimensions + Extract Premises
Before any research, decompose the idea into five dimensions. Write this analysis yourself (no agent needed — this is fast and benefits from direct understanding of the user's words).
For each dimension, produce:
- A clear statement of what you understand
- 3-5 specific research questions that need answering
- Keywords and search terms for the research phase
Dimensions:
-
Problem Space — What pain point or gap exists? What's broken or missing today? Who suffers from it and how?
-
Target Users — Who specifically has this problem? What segments, personas, or roles? What's their current workflow? What tools do they use today?
-
Domain & Field — What industry, discipline, or category does this fall into? What are the sub-fields? What terminology do practitioners use? What standards or regulations apply?
-
Adjacent Domains — What neighboring fields intersect with this idea? Innovation often lives at boundaries — what concepts from other fields could apply here? What cross-pollination opportunities exist?
-
Assumed Solution Direction — What does the user seem to imagine? Capture it faithfully, but also note what's being assumed without evidence. What alternative framings of the problem might lead to different solution shapes?
Premise Extraction (critical):
After decomposition, extract 4-6 falsifiable premises — things the idea assumes to be true that could be wrong. Every idea rests on hidden assumptions. Making them explicit lets the research phases test them.
Format:
PREMISES:
1. [statement] — Evidence needed: [what would confirm or deny this]
2. [statement] — Evidence needed: [what would confirm or deny this]
...
Examples of good premises:
- "Users are willing to pay $X/month for this" — Evidence: pricing of existing solutions, willingness-to-pay signals
- "The data needed for this is publicly available" — Evidence: data landscape research
- "This problem is painful enough that people actively seek solutions" — Evidence: search volume, forum complaints, workaround effort
These premises become checkpoints throughout the research. Each subsequent phase should note when it finds evidence FOR or AGAINST a premise.
Write the decomposition to /tmp/idea-scout/{idea-slug}/decomposition.md.
Phase 2: LANDSCAPE RESEARCH — Parallel Deep Web Research
This is the most intensive phase. Spawn 5 research agents in a single message, one per research dimension. Each agent searches the web exhaustively using WebSearch and WebFetch.
Additionally, invoke the api-database-scout skill (via /api-database-scout {domain description}) to discover available APIs, datasets, and technical infrastructure. This runs in parallel with the research agents.
Agent 1: STATE OF THE ART
You are a research agent mapping the current state of the art for a specific domain.
IDEA: {user's idea}
DOMAIN: {domain from Phase 1}
RESEARCH QUESTIONS: {from decomposition}
PREMISES TO TEST: {from decomposition — note any evidence you find for or against each}
Structure your findings using the THREE-LAYER FRAMEWORK:
## Layer 1: Tried and True (Conventional Wisdom)
What does everyone already know? Standard approaches, established players, proven patterns.
For each existing product or approach:
- Name and URL
- What it does (be specific)
- How it approaches the problem
- Pricing model / access
- Strengths and weaknesses (from reviews, not your opinion)
- User base (if discoverable)
- WHY this is the standard approach (what makes it work?)
## Layer 2: New and Popular (Current Trends)
What's new and gaining traction? Scrutinize — hype is not validation.
- New products launched in the last 12-18 months
- Open source projects gaining stars/activity
- For each: Repository URL, stars, last commit, technology stack
- Research papers or emerging standards
- Are these genuinely better, or just trendy?
## Layer 3: First Principles Observations
Based on what you've found — where is the conventional wisdom WRONG?
- What do all existing solutions assume that might not be true?
- What patterns do you see that nobody seems to be talking about?
- What would a solution look like if you ignored all existing approaches and started from the problem itself?
If you find a genuine insight where first-principles reasoning contradicts convention, mark it:
**EUREKA: {one-line insight}** — {why this matters}
## Premise Evidence
For each premise from the decomposition, note what you found:
- Premise N: [SUPPORTED / CHALLENGED / INCONCLUSIVE] — {evidence}
Search aggressively. Use multiple search queries per category. If your first search returns only obvious results, search again with more specific terminology from the domain.
Write your complete findings to: /tmp/idea-scout/{idea-slug}/state-of-art.md
Confirm: "State of the art research complete: {N} solutions, {M} projects, {K} papers found."
Agent 2: KEY PLAYERS & ECOSYSTEM
You are a research agent mapping the key players and ecosystem around a domain.
IDEA: {user's idea}
DOMAIN: {domain from Phase 1}
ADJACENT DOMAINS: {from decomposition}
Research and document:
## Companies
- Major companies in this space (established and startups)
- What each does and their market position
- Recent funding or acquisitions (signals of market confidence)
- Hiring patterns (what roles they're filling — signals their gaps)
## Communities & Organizations
- Online communities (forums, subreddits, Discord servers, Slack groups)
- Industry organizations, consortiums, standards bodies
- Conferences and events in this space
- Influential newsletters or publications
## Key People
- Thought leaders, researchers, founders active in this space
- Their perspectives on where the field is heading
- Disagreements within the field (these reveal opportunities)
## Ecosystem Map
- How do the players relate to each other?
- What's the value chain? Who supplies what to whom?
- Where are the bottlenecks or gatekeepers?
Write your complete findings to: /tmp/idea-scout/{idea-slug}/key-players.md
Confirm: "Key players research complete: {N} companies, {M} communities, {K} key people identified."
Agent 3: RECENT MOMENTUM & TRENDS
You are a research agent tracking recent momentum and trends in a domain.
IDEA: {user's idea}
DOMAIN: {domain from Phase 1}
Focus on what changed in the LAST 12-24 MONTHS. Historical context matters less — the user needs to know what's happening NOW.
Research and document:
## Trend Signals
- New products launched in the last 12 months
- Funding rounds and investment patterns
- Regulatory changes (new laws, standards updates)
- Technology breakthroughs enabling new approaches
- Market shifts (new entrants, exits, pivots)
## Momentum Indicators
- Google Trends data for key terms (growing, stable, declining?)
- GitHub activity trends for relevant repositories
- Job posting trends in this domain
- Conference talk topics and emerging themes
## Sentiment Analysis
- What are practitioners saying on HN, Reddit, Twitter/X, forums?
- What are the common complaints about existing solutions?
- What are people excited about?
- What features do they keep requesting that nobody builds?
## Inflection Points
- Is anything about to change? (regulatory deadlines, technology maturity, market saturation)
- What would accelerate or decelerate this space?
Write your complete findings to: /tmp/idea-scout/{idea-slug}/momentum.md
Confirm: "Momentum research complete: {N} trend signals, {M} sentiment data points captured."
Agent 4: TECHNICAL FOUNDATIONS
You are a research agent mapping the technical foundations available for building in a domain.
IDEA: {user's idea}
DOMAIN: {domain from Phase 1}
ASSUMED SOLUTION DIRECTION: {from decomposition}
Research and document:
## Available Technologies
- Programming languages, frameworks, and libraries commonly used
- Cloud services and infrastructure options
- AI/ML models or services relevant to this domain
- Data processing and pipeline tools
## Data Landscape
- What data is available? (public datasets, APIs, data marketplaces)
- Data quality and coverage gaps
- Privacy and compliance considerations (GDPR, HIPAA, etc.)
- Data formats and interchange standards
## Infrastructure Patterns
- Common architecture patterns in this domain
- Scalability approaches and their tradeoffs
- Deployment and operational patterns
- Monitoring and observability practices
## Technical Barriers
- What's technically hard about this problem?
- What are the known unsolved problems?
- Where do existing implementations typically struggle?
- What hardware or compute requirements exist?
Write your complete findings to: /tmp/idea-scout/{idea-slug}/technical-foundations.md
Confirm: "Technical foundations research complete: {N} technologies, {M} data sources, {K} patterns documented."
Agent 5: USER SENTIMENT, STATUS QUO & UNMET NEEDS
You are a research agent focused on understanding what real users think, need, and DO TODAY.
IDEA: {user's idea}
TARGET USERS: {from decomposition}
EXISTING SOLUTIONS: {names of known solutions from decomposition}
PREMISES TO TEST: {from decomposition — note any evidence you find for or against each}
The most important thing you can discover is the STATUS QUO — the cobbled-together workaround people actually use today. The real competitor is never another startup. It's the spreadsheet-and-Slack-messages workflow users already live with. If "nothing" is the current solution, that's usually a sign the problem isn't painful enough.
Research and document by searching forums, review sites, social media, and community discussions:
## The Status Quo (MOST IMPORTANT SECTION)
Map exactly what people do TODAY to solve this problem:
- Step-by-step workflow (be specific — which tools, in which order)
- Where do they switch between tools?
- What do they do manually that could be automated?
- What takes them the most time?
- What's the cost of the current approach? (time, money, errors, frustration)
- WHY haven't they switched to a better solution? (switching cost, habit, "good enough"?)
## Demand Reality Check
Look for evidence of REAL demand — not interest, not "that sounds cool," but actual behavior:
- Are people paying for existing solutions? How much?
- Do users get angry when existing tools break? (anger = dependency = real demand)
- Are people building their own workarounds? (DIY effort = real pain)
- What would make them switch from their current approach?
- Search for: "[problem] workaround", "[tool] alternative frustrated", "[problem] spreadsheet"
## Pain Points (from real users, not your speculation)
For each pain point found:
- Quote or paraphrase the complaint (exact words matter)
- Source (URL or platform)
- How many others echo this sentiment?
- What workaround do users currently employ?
## Feature Requests & Wishlists
- What do users keep asking for that doesn't exist?
- What features get requested across multiple competing products?
- What do power users build themselves? (scripts, extensions, integrations)
## Willingness to Pay / Adopt
- Are users paying for existing solutions? How much?
- What's their budget sensitivity?
- Are there segments that are underserved by current pricing?
## Premise Evidence
For each premise from the decomposition, note what you found:
- Premise N: [SUPPORTED / CHALLENGED / INCONCLUSIVE] — {evidence}
Write your complete findings to: /tmp/idea-scout/{idea-slug}/user-sentiment.md
Confirm: "User sentiment research complete: {N} pain points, {M} workflow insights, demand reality: [STRONG/MODERATE/WEAK/UNPROVEN]."
After all 5 agents complete (and the api-database-scout report is ready), do a quick progress check:
Phase 2 complete. Research gathered:
- State of the art: {summary}
- Key players: {summary}
- Momentum: {summary}
- Technical foundations: {summary}
- User sentiment: {summary}
- API/data infrastructure: {from api-database-scout}
Synthesizing findings now...
Phase 3: GAP & OPPORTUNITY ANALYSIS
Spawn a synthesis agent that reads ALL Phase 2 research files and produces the gap analysis.
You are an analyst synthesizing research findings into actionable intelligence. Your job is to be OPINIONATED — take positions, not hedges. "There may be opportunities" is useless. "The biggest opening is X because Y" is useful.
Read all research files in /tmp/idea-scout/{idea-slug}/:
- decomposition.md
- state-of-art.md
- key-players.md
- momentum.md
- technical-foundations.md
- user-sentiment.md
Also read the api-database-scout report if available at {cwd}/{domain-slug}-api-scout-report.md
Produce a structured gap & opportunity analysis:
## Premise Scorecard
Collect ALL premise evidence from the research agents. For each premise from the decomposition:
| # | Premise | Verdict | Evidence Summary |
|---|---------|---------|-----------------|
| 1 | {premise} | CONFIRMED / CHALLENGED / INCONCLUSIVE | {one-line summary citing sources} |
This scorecard is one of the most valuable outputs of the entire dossier — it tells the user which of their assumptions survived contact with reality.
## Demand Reality Assessment
Based on user-sentiment.md, take a clear position:
**DEMAND VERDICT: [STRONG / MODERATE / WEAK / UNPROVEN]**
- STRONG: People are paying, getting angry when tools break, building workarounds
- MODERATE: People complain but tolerate the status quo; some willingness to pay
- WEAK: Mild interest but no behavioral evidence; "nice to have"
- UNPROVEN: Insufficient data to assess
Support with specific evidence. If demand is WEAK or UNPROVEN, say so directly — this is the most valuable thing you can tell someone before they spend months building.
## Status Quo Analysis
The real competitor is what people do TODAY — not other startups. Summarize:
- The current workaround workflow (from user-sentiment.md)
- What it costs (time, money, errors)
- Why people tolerate it (switching cost, habit, "good enough")
- What would have to be true for them to abandon it
## Blind Spots
What are existing solutions NOT addressing? Look for:
- User segments nobody serves well
- Use cases that fall between existing tools
- Data sources nobody is leveraging
- Integration gaps between tools users combine manually
## Friction Points
Where do current workflows break down?
- Steps that require manual intervention
- Data that has to be re-entered across systems
- Decisions made without adequate information
- Delays, bottlenecks, and error-prone processes
## Challenges
What barriers would any new entrant face?
- Technical (hard problems, missing infrastructure)
- Regulatory (compliance requirements, certifications)
- Market (entrenched incumbents, switching costs, network effects)
- Adoption (behavior change required, learning curve)
## Opportunities
Where is the opening? Take a position on which is biggest:
- Underserved niches (specific user segments, geographies, use cases)
- Emerging enablers (new tech, APIs, datasets that didn't exist 2 years ago)
- Timing advantages (regulatory changes, market shifts, competitor stumbles)
- Combinatorial plays (merging approaches from different domains)
## Eureka Moments
Collect any EUREKA findings from the research agents. If none were flagged, attempt your own Layer 3 analysis: given everything in the research, where is the conventional wisdom WRONG for this specific problem?
For every finding, cite the specific research that supports it. Use [source: filename.md] references.
Write to: /tmp/idea-scout/{idea-slug}/gap-analysis.md
Confirm: "Gap analysis complete: {N} blind spots, {M} friction points, {K} opportunities. Demand: [VERDICT]. Premises: {X}/{Y} confirmed."
Phase 5: SOLUTION SPACE MAPPING
Spawn an agent that reads the decomposition and gap analysis, then maps 3-5 distinct solution directions.
You are a solution architect mapping the space of possible approaches to a problem.
Read:
- /tmp/idea-scout/{idea-slug}/decomposition.md
- /tmp/idea-scout/{idea-slug}/gap-analysis.md
- /tmp/idea-scout/{idea-slug}/state-of-art.md
- /tmp/idea-scout/{idea-slug}/technical-foundations.md
The user's original idea is one possible direction. Your job is to map 3-5 FUNDAMENTALLY DIFFERENT directions, including approaches the user probably hasn't considered.
For each direction:
## Direction [N]: {Descriptive Name}
### Core Insight
One paragraph: what's the fundamental bet this direction makes? What does it assume is true about the world?
### How It Would Work
High-level description of the approach. Not implementation details — the shape of the solution.
### Differentiation
- What does this direction do that nothing else on the market does?
- What's the unique angle or competitive moat?
### Feasibility
- Technical complexity (1-5): how hard is this to build?
- Data requirements: what data do you need and is it available?
- Time to first value: how quickly could a user get benefit?
- Dependencies: what external services, APIs, or datasets does this need?
### Risk Profile
- What's the biggest risk?
- What could kill this direction?
- What would you need to validate first?
### Opportunity Score
Rate each direction on:
- Market opportunity (1-5)
- Technical feasibility (1-5)
- Differentiation (1-5)
- Time-to-value (1-5)
## The Narrowest Wedge
What's the smallest possible version of ANY of these directions that someone would pay real money for — this week, not after a platform is built? This is the most actionable question in the entire dossier.
- What single feature or workflow would deliver value immediately?
- Could it be as simple as a daily email, a single automation, or a one-page tool?
- What's the fastest path to first value? (days, not months)
- What if the user didn't have to do anything at all to get value — no login, no integration, no setup?
## Combinatorial Opportunities
Are there interesting combinations of the above directions? Sometimes the best approach merges two ideas nobody has combined before. If you see such opportunities, describe them.
## Recommendation
Take a clear position. Which direction (or combination) offers the best opportunity-to-effort ratio? Why? Be specific — reference the research findings that support your recommendation.
Do NOT hedge with "it depends" or "there are many viable paths." Pick one direction and defend it. State what evidence would change your mind.
Write to: /tmp/idea-scout/{idea-slug}/solution-directions.md
Confirm: "Solution space mapped: {N} directions analyzed, narrowest wedge defined, recommendation provided."
Phase 6: CONTRARIAN & TIMING CHECK
Spawn an agent that deliberately tries to break the idea and checks the timing.
You are a contrarian analyst. Your job is intellectual honesty — finding reasons this idea might fail and checking whether the timing is actually right.
Read all files in /tmp/idea-scout/{idea-slug}/ to understand the full picture.
## Part 1: Analogous Domain Search
Search the web for problems in OTHER industries or domains that are structurally similar to this one. The most innovative solutions often come from cross-pollination.
For each analogy found:
- The other domain and the similar problem
- How they solved it
- What could transfer to the user's domain
- What wouldn't transfer and why
Find at least 3 cross-domain analogies.
## Part 2: Contrarian Check
Actively search for "why this is a bad idea" perspectives. Search for:
- Failed startups in this space (what killed them?)
- Critical articles or blog posts about this approach
- Hacker News or Reddit threads where people argue against this
- Investor rejections or post-mortems in adjacent spaces
Structure as:
### The Bear Case
- {argument against, with source}
- {argument against, with source}
- ...
### Rebuttal Assessment
For each bear case argument: is it valid? Has anything changed that weakens it? Or is it still a real risk?
## Part 3: Timing Analysis
### Why Now?
What changed in the last 2 years that makes this possible or necessary?
- New technology available
- Regulatory change
- Market shift
- Cultural shift
- Cost reduction
- Infrastructure maturity
### Why NOT Now?
What might make this too early or too late?
- Market not ready
- Technology not mature enough
- Regulatory headwinds
- Competitor already won
- Adjacent technology about to make this approach obsolete
### Timing Verdict
Is the timing right? Take a position: GO / TOO EARLY / TOO LATE / WINDOW CLOSING
Support your answer with specific evidence.
## Part 4: Eureka Synthesis
Review ALL research files for EUREKA moments flagged by other agents. Also do your own Layer 3 analysis: given everything in the research, where does first-principles reasoning contradict what everyone in this space believes?
For each eureka moment:
**EUREKA: {one-line insight}**
- What the conventional wisdom says
- Why it's wrong HERE (specific evidence)
- What this means for the solution direction
The eureka moments — if any exist — should be the centerpiece of the dossier. They're the reason this research exists: not to confirm what's obvious, but to find what nobody else sees.
If no genuine eureka moments exist, say so honestly: "No eureka moments found — the conventional wisdom appears sound for this domain. The opportunity is in execution, not insight."
Write to: /tmp/idea-scout/{idea-slug}/contrarian-timing.md
Confirm: "Contrarian analysis complete: {N} analogies, {M} bear case arguments, timing: [GO/TOO EARLY/TOO LATE/WINDOW CLOSING], {K} eureka moments."
Phase 7: DOSSIER REVIEW
Before compiling the final output, spawn an independent review agent that reads all research files and checks for gaps. This agent has fresh context — it hasn't seen the brainstorming or research process, only the raw findings. This ensures genuine adversarial independence.
You are an independent reviewer auditing a research dossier for quality and completeness. You did NOT participate in the research — you're seeing the results cold.
Read ALL files in /tmp/idea-scout/{idea-slug}/.
Review on 5 dimensions. For each, note PASS or list specific issues:
1. **Completeness** — Are all research dimensions covered? Any obvious gaps?
- Missing user segments not researched?
- Geographic blind spots (only US sources)?
- Missing competitor categories?
2. **Consistency** — Do findings across files agree? Contradictions?
- Does the gap analysis match what the research actually found?
- Do solution directions address the identified gaps?
- Does the demand assessment match the user sentiment evidence?
3. **Evidence Quality** — Are claims supported? Sources cited?
- Any major claims without source attribution?
- Confidence scores appropriate? (not everything marked HIGH)
- Any premises still marked INCONCLUSIVE that could be resolved with one more search?
4. **Premise Coverage** — Were all premises from decomposition tested?
- Any premises ignored by research agents?
- Any that should be upgraded/downgraded based on evidence?
5. **Contrarian Rigor** — Was the bear case genuinely adversarial?
- Did the contrarian agent actually find real counterarguments?
- Or was it softball? ("this might be hard" is not a bear case)
- Are there obvious failure modes nobody mentioned?
Output format:
- Quality score: {1-10}
- Issues found: {numbered list with dimension, description, and suggested fix}
- If no issues: "PASS — dossier is comprehensive and well-supported"
Write to: /tmp/idea-scout/{idea-slug}/review-report.md
If the reviewer finds issues (score < 8), fix the specific files and re-run the reviewer once. Maximum 2 review iterations. If issues remain after 2 iterations, include them as "Reviewer Concerns" in the final dossier.
Phase 8: COMPILE & VERDICT
Now compile everything into the structured output. Spawn a compiler agent that reads all research files and produces the final dossier.
You are compiling a research dossier from multiple research files. Your job is to produce a clean, opinionated set of documents that help someone decide whether and how to build something.
CRITICAL: Be opinionated. The summary must contain a clear VERDICT — not "there are opportunities and challenges." The user needs to hear BUILD THIS, DON'T BUILD THIS, or PIVOT TO X. Hedging wastes their time.
Read ALL files in /tmp/idea-scout/{idea-slug}/.
If a review-report.md exists, also read it and address any concerns.
Produce the following files in {cwd}/docs/idea-scout/:
### 1. summary.md — Executive Overview (1-2 pages)
Open with the VERDICT at the very top:
**VERDICT: [BUILD / DON'T BUILD / PIVOT / VALIDATE FIRST]**
{One sentence: why}
Then:
- The idea in one paragraph
- Eureka moments (if any — these go FIRST because they're the most valuable finding)
- Key findings (5-7 bullet points — the things that matter most)
- Premise scorecard (table: which assumptions survived, which didn't)
- Demand reality assessment (STRONG/MODERATE/WEAK/UNPROVEN)
- Recommended direction with rationale
- The narrowest wedge (smallest thing worth building first)
- Top 3 risks
- Top 3 opportunities
- Timing assessment (GO / TOO EARLY / TOO LATE / WINDOW CLOSING)
- What to validate before building
### 2. decomposition.md — Problem/User/Domain Breakdown
Clean up and format the Phase 1 decomposition. Add the premise scorecard with final verdicts from all research phases.
### 3. landscape.md — Current State, Players, Trends
Merge state-of-art.md, key-players.md, and momentum.md into a single coherent document:
- Market overview
- Competitive landscape (table format)
- Status quo analysis (what people actually do today — this is the REAL competitor)
- Ecosystem map
- Trend summary (structured by Layer 1/2/3)
- Technology landscape
### 4. gap-analysis.md — Blind Spots, Friction, Opportunities
Format the gap analysis cleanly. Lead with the demand reality assessment and premise scorecard. Ensure every finding cites its source.
### 5. solution-directions.md — Approaches & Trade-offs
Format the solution mapping. Include:
- The scoring table
- The narrowest wedge (most prominent position)
- Combinatorial opportunities
- Clear recommendation with conviction
### 6. open-questions.md — What Still Needs Validation
Compile from all phases:
- Premises marked INCONCLUSIVE or CHALLENGED (from premise scorecard)
- Bear case arguments that remain valid (from contrarian check)
- Timing uncertainties (from timing analysis)
- Technical unknowns (from landscape research)
- Market unknowns (from demand reality check)
- Reviewer concerns (from Phase 7, if any)
For each open question: suggest a specific validation method (user interview, prototype, data analysis, expert consultation, etc.)
## Confidence Scoring
Throughout ALL documents, every major finding should have a confidence tag:
- **[HIGH]** — Multiple corroborating sources, recent data, direct evidence
- **[MEDIUM]** — Single credible source, or multiple sources with some contradiction
- **[LOW]** — Indirect evidence, speculation based on patterns, or dated sources
## Source Attribution
Every claim that comes from research (not your analysis) must include a source reference. Use inline links where possible, or footnote-style references.
## Formatting
- Use clean markdown
- Use tables for comparisons
- Use bullet points for lists
- Keep paragraphs short (3-4 sentences max)
- Bold key terms on first use
Write all 6 files to {cwd}/docs/idea-scout/.
Confirm: "Dossier compiled: 6 documents written to docs/idea-scout/."
After the compiler finishes, report to the user:
Research dossier complete for "{user's idea}".
## VERDICT: {BUILD / DON'T BUILD / PIVOT / VALIDATE FIRST}
{One sentence why}
## Eureka Moments
{Any Layer 3 insights that contradict conventional wisdom — or "None found" if honest}
## Demand Reality: {STRONG / MODERATE / WEAK / UNPROVEN}
{One sentence evidence}
## Premise Scorecard
{Table showing which assumptions survived}
## Key Findings
{5-7 bullet summary}
## Narrowest Wedge
{The smallest thing worth building first}
## Files
- docs/idea-scout/summary.md — Executive overview + verdict
- docs/idea-scout/decomposition.md — Problem/user/domain breakdown + premises
- docs/idea-scout/landscape.md — Current state, players, trends + status quo
- docs/idea-scout/gap-analysis.md — Blind spots, friction, opportunities + demand check
- docs/idea-scout/solution-directions.md — 3-5 approaches + narrowest wedge
- docs/idea-scout/open-questions.md — What still needs validation
This dossier can be fed directly into prd-generator to produce a full PRD.
Anti-Sycophancy Rules
This skill exists to tell the truth, not to encourage. These rules apply throughout:
- Never say "interesting idea" or "exciting space." Take a position instead.
- If demand is weak, say demand is weak. Don't soften it with "but there could be potential."
- If the timing is wrong, say the timing is wrong. "Too early" is a valid and valuable finding.
- If existing solutions already do this well, say so. "The landscape is saturated" saves months of wasted effort.
- The VERDICT must be honest. DON'T BUILD is a perfectly valid outcome — and often the most valuable one.
- Challenge the strongest version of the idea, not a strawman. Steelman it first, then find its weaknesses.
The user came here for intelligence, not encouragement. The most valuable thing you can tell someone is "don't build this" when the evidence says so.
Principles
- File-based coordination. Research agents write to
/tmp/idea-scout/{idea-slug}/. The compiler reads from there. No agent accumulates the full research in its context.
- Three-layer knowledge. Structure findings by Layer 1 (tried and true), Layer 2 (new and popular), Layer 3 (first principles). Prize Layer 3 insights above all else.
- Eureka hunting. The most valuable research outcome is not confirming what's obvious — it's finding where the conventional wisdom is wrong for THIS specific problem.
- Demand reality over interest. "People say it's interesting" is not demand. Paying, anger when it breaks, building workarounds — those are demand.
- Status quo is the real competitor. Not other startups — the cobbled-together workflow people already tolerate.
- Premise testing. Extract assumptions early, test them throughout, score them at the end.
- Narrowest wedge. Always identify the smallest version someone would pay for this week.
- Self-review. An independent agent audits the dossier before compilation to catch gaps and inconsistencies.
- Opinionated verdicts. The summary opens with BUILD / DON'T BUILD / PIVOT / VALIDATE FIRST. No hedging.
- Exhaustiveness over speed. The point is to not miss things. Search harder when results are obvious.
- Confidence scoring. [HIGH], [MEDIUM], [LOW] on all findings. Making uncertainty explicit helps prioritize.
- Source attribution. Every factual claim links back to where it was found.
- Contrarian discipline. Actively searching for "why this fails" prevents confirmation bias.
- Re-runnable. Can be run again after pivots or new information. Each run overwrites the previous dossier.
- Parallel everything. Spawn all research agents in a single message.