competitor-recon 竞品调研

Run competitor reconnaissance before building any feature: the review sections of existing products are the cheapest requirements research that exists.

动手做功能之前先跑竞品侦察：现有产品的用户评论区，是世界上最便宜的需求调研。

Why / 为什么

Marketing pages state intentions; user reviews state outcomes. One grounded complaint is worth more than any amount of feature-page copy — and a complaint repeated by three independent users is an unmet need you can build against.

营销页写的是意图，用户评论写的是结果。一条有出处的真实抱怨，胜过整页功能文案；三个互不相识的用户抱怨同一件事，就是一个可以动手满足的未满足需求。

The method / 方法（4 步）

1. Enumerate 3–6 competitors — search English and Chinese sources; one language samples half the market.
2. Extract per competitor: features, pricing (numbers, not "contact us"), positioning, and user complaints with source links.
3. Cluster complaints by the underlying need, ranked by frequency.
4. Deliver: comparison table + copy-worthy list + skip list. A feature idea enters the build list only with a cited user complaint behind it.

Worked example — recon on an open-source ledger app (2026-09, real run)

Sources actually fetched during the run: V2EX threads (via curl), the target's GitHub issue tracker (via API — reaction counts double as frequency statistics), community forums.

Complaint cluster	Evidence	Verdict
Ads / subscription inflation driving users out	multiple independent forum posts naming the incumbents	differentiate: stay clean, price predictably
"AI auto-capture might be wrong, I re-check manually"	feature request with 👍×3 in the target's own tracker	build: recognition → confirmation loop
Reimbursement / refund flows too complex	complaints across two products, praise for a third's	build: expense-return workflow done simply
Multi-currency locked behind a paywall	a long-time user defected over exactly this	skip: no demand evidence beyond the wall itself

The run produced the full comparison table plus a skip list with reasons — 3 build ideas, each cited; 3 skips, each justified.

Rules that make it work / 让它生效的规则

• Reviews outrank marketing pages; conflicts are recorded as claims.
• Complaint clusters are the finding — one is anecdote, three is a need.
• Issue trackers are goldmines: a 👍 count is a frequency statistic for free.
• Tool blocked? Switch clients, not sources (curl, public APIs). The evidence bar never drops.

Honest limitations / 如实说明局限

• Review sites and forums rate-limit bots; some evidence may come from search snippets instead of full threads — those citations are marked as such.
• Reaction counts measure engagement, not market size; a quiet competitor may simply have no community.
• English+Chinese covers our ecosystems; other markets (JP/KR/EU) are not sampled.

评论站对爬虫限流；部分证据可能来自搜索摘要而非完整帖子——这类引用会明确标注。反应数衡量参与度而非市场规模。我们只覆盖中英两个生态。

Install / 安装

git clone https://github.com/ChenneyZhuang/competitor-recon ~/.claude/skills/competitor-recon

One SKILL.md. Needs only web search and any fetch client. MIT. v0.2.0 — live-tested end-to-end.

单个 SKILL.md，只需搜索与抓取能力。MIT 许可，v0.2.0，完整实测通过。