A Claude Code plugin that validates whether a demand is real — before you build it. Frame a falsifiable hypothesis, fan out parallel agents across two evidence pillars, score it on 7 axes with source triangulation, run an adversarial red-team that tries to kill your idea, and ship an answer-first HTML report with a clear verdict: Go / Conditional / Pivot / No-go. Zero login, no scrapers.
🌏 中文版 · 📖 Methodology · 🖥 Live page · ⚖️ License
中文简介： 一款 Claude Code 插件，在你动手做之前帮你验证「这个需求到底是不是真的」。把模糊想法逼成可证伪假设，并行派多个 agent 从两根证据支柱采集，套 7 项评分卡 + 来源三角验证，再让一个对抗红队专门证伪，最后产出一份「答案优先」的 HTML 报告，给出明确裁决：可以做 / 有条件做 / 转向 / 别做。全程零登录、不写爬虫。完整中文文档 → README_zh.md
You have a product idea. Before writing a line of code, you want to know: is this demand real, how big is it, who has it, what do they use today, and will they pay? Most founders either guess, or drown in scattered Google tabs and confirmation bias.
The tools that exist are either single-source (Reddit-only pain miners) or generic "deep research" with no validation framework and no devil's advocate. Demand Radar packages multi-source evidence collection + a demand scorecard + an adversarial red-team into one repeatable run — and it's built to fight your confirmation bias, not feed it.
🎯 Real evidence, two pillars. Every verdict stands on both top-down (industry reports, market size, growth, capital signals) and bottom-up (Hacker News, App Store / Google Play reviews, search demand, social). When the two disagree — "analysts say the market is huge but nobody's complaining" — that contradiction is the finding.
🔴 An adversarial red-team, by design. You have confirmation bias; the red-team doesn't. A dedicated agent's only KPI is to kill your hypothesis — fake demand, saturated, declining, nobody pays, can't reach them, sample bias. Any "Go" has to survive it. In testing it pulled a verdict from Conditional (55%) down to Pivot (40%) by surfacing a buyer-mismatch the optimistic read had missed.
📐 A 7-axis scorecard with hard rules. Pain intensity · prevalence · current-alternative gap · willingness to pay (weighted ×2) · market size & trend · differentiation wedge · reachability. Competition counts as proof of demand, not an automatic penalty — a red ocean only costs you the wedge axis when there's no opening left to take. A signal counts as verified only when ≥2 independent sources confirm it — single-source stays a "lead." Three independent judge agents score in parallel; disagreements are flagged, not averaged away.
📋 Answer-first reports, not analyst dumps. The report opens with your question → a plain-language verdict → why — the score sits on a 0–100 band so you can see at a glance what it means. Scorecard, what-to-do, details, and the full method/evidence trail follow, with the evidence collapsed. Every claim cites an evidence ID you can click to trace.
🔒 Zero login, no scrapers, no keys. Official free APIs (Hacker News, iTunes; Google Play optional, needs a pip library) + web search only. Nothing to register, no OAuth, no anti-bot games. Reports render to a single self-contained HTML file.
🌐 Speaks your language. Ask in English → English report. 用中文问 → 中文报告.
中文要点：
- 🎯 双支柱证据：自上而下（行业报告/市场规模/资本）+ 自下而上（HN/应用商店评论/搜索/社交），上下矛盾本身就是关键发现
- 🔴 内置对抗红队：唯一 KPI 是杀死你的假设，任何「可以做」都要先扛住它（实测把「有条件做 55%」拉到「转向 40%」）
- 📐 7 项评分卡 + 硬规则：付费意愿 ×2；信号需 ≥2 独立源才算「已验证」；3 评委并行打分，分歧标红不和稀泥
- 📋 答案优先：报告开头就是「你的问题 → 大白话裁决 → 该怎么做」，分析过程和证据折叠在后面，每条结论可点击溯源
- 🔒 零登录、不写爬虫：只用免费官方 API + 网页搜索，无需注册/OAuth/key
- 🌐 跟你的语言走：英文问出英文报告，中文问出中文报告
完整中文版 → README_zh.md
Run /demand <your idea> and get a single-file, answer-first HTML report:
- Verdict — Go / Conditional / Pivot / No-go, with a plain-language answer to "is my demand real, and why," demand-state tags (real demand / red ocean / niche wedge…), a confidence level, and the score on a 0–100 band that shows what it means.
- Scorecard — 7 axes with visual bars, post-red-team final scores, and the weighted total.
- What to do next — concrete, pastable action cards (change your buyer, narrow the niche, run a pre-sale), prioritized. The part founders actually use.
- Details — what each source domain says, the two-pillar reconciliation, the red-team's most lethal flaw, a current-alternatives map, and an explicit "what we could NOT verify" section. No over-promising.
- Method — collapsed at the bottom: agent count, evidence count, the full pipeline, and a clickable evidence appendix.
Step 1 — Add the marketplace:
/plugin marketplace add LeifDiao/demand-radar
Step 2 — Install the plugin:
/plugin install demand-radar@demand-radar-marketplace
Alternative (local):
git clone https://github.com/LeifDiao/demand-radar.git ~/demand-radar
claude --plugin-dir ~/demand-radar/demand a simpler Notion for small remote teams
- Demand Radar turns your idea into a falsifiable hypothesis (ICP / job / current workaround / the claim / Go criteria) — it'll ask if anything's unclear.
- It fans out parallel collector agents across both pillars (real searches + free APIs).
- It triangulates evidence, runs 3 independent judges, then the red-team.
- The report opens in your browser.
- Claude Code with plugin support
- Python 3 (for the connectors and report generator — standard library only)
- Optional:
pip install google-play-scraperto enable Google Play reviews - No API key, no build step, no server
Two-layer model — the model judges, scripts enforce:
- Evidence — parallel agents collect from official free APIs + web search, each item tagged with a
signaland a source type (hard / behavioral / opinion / analyst).triangulate.pyenforces the ≥2-independent-sources rule deterministically. - Scoring — three judge agents (neutral / optimistic / strict) score the 7 axes independently;
aggregate_scores.pytakes the median, flags disagreements, and computes the weighted verdict. The red-team then re-scores in a closed loop.
Anti-bias mechanisms:
- Adversarial red-team before any verdict is final
- Triangulation — single-source signals can't masquerade as verified
- Forced "what we couldn't verify" section — "not found" is never written as "doesn't exist"
- Opinion ≤ lead — "I'd buy it" never supports a willingness-to-pay conclusion on its own
demand-radar/
├── .claude-plugin/
│ ├── plugin.json # plugin manifest
│ └── marketplace.json # marketplace entry
├── skills/demand/ # invoked as /demand-radar:demand
│ ├── SKILL.md # frame → fan-out → triangulate → judge → red-team → report
│ ├── references/ # validation framework, source playbook, red-team checklist, templates
│ └── scripts/
│ ├── connectors/ # hn_algolia · itunes · play_reviews (free APIs) · reddit (best-effort)
│ ├── triangulate.py # ≥2-source confidence rule
│ ├── aggregate_scores.py # multi-judge median + disagreement flags + weighted verdict
│ └── generate_report.py # report.json → answer-first HTML
└── docs/
├── index.html # landing page
├── sample-report.html # example report
└── METHODOLOGY.md / METHODOLOGY_zh.md
Zero runtime dependencies (Google Play reviews optional).
Demand Radar is released under CC BY-NC 4.0:
- ✅ Free for personal, educational, research, and any non-commercial use
- ✅ Forking, modifying, sharing is welcome — please attribute the original repo and indicate changes
- ❌ Commercial use (bundling into paid products, paid SaaS hosting, selling reports based on the scoring) requires a separate license
For commercial licensing, contact: leifdiao@gmail.com
See LICENSE for full terms, including the 中文版说明.
Built for people who want to know if the demand is real — before they build it.