360 Secure SkillHub
360 安全技能中心

Pain to MVP

该 Skill 提供发现与 PRD 两种模式:发现模式借助外部搜索/抓取工具收集公开讨论证据,经 npm 脚本写入 TiDB 后聚类排序产品机会;PRD 模式基于已持久化的机会与证据生成轻量 PRD。强调证据可追溯、不虚构 SQL、证据薄弱时明示证据不足。

zhao-weijie
zhao-weijie
热度 44.35
最近更新 2026-05-11 17:37:59
分类:营销依赖:代码执行 / 读写文件 / 内部调用大模型 / 第三方 API应用场景:市场分析 / 研究 / 报告撰写 / 工作流 ...

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add pain2mvp
Metadata
name
pain2mvp
description
Discover and structure product opportunities from public user discussions, then convert top opportunities into a lightweight PRD for a coding agent. Use when the user needs direction on what to build, or wants to find recurring pain points, unmet needs, competitor complaints, or market signals around an idea, user group, or product category.
version
1.0.2
metadata
{"openclaw":{"install":[{"kind":"brew","formula":"jq","bins":["jq"]}],"skillKey":"pain2mvp","requires":{"bins":["node","npm"],"env":["TIDB_DATABASE_URL"]},"primaryEnv":"TIDB_DATABASE_URL"}}

Painpoint To MVP

Modes

  1. Discovery mode: find and rank product opportunities from public user discussion evidence.
  2. PRD mode: turn one persisted opportunity into a lightweight PRD for a coding agent.

If the user asks for both, discovery comes first.

Hard Rules

  • references/contracts.md is the only runtime contract reference.
  • Only inspect scripts when a documented helper command fails unexpectedly or when explicitly modifying this skill.
  • Read references/contracts.md once per run, not repeatedly.

Preferred Commands

Use these commands for normal operation:

  • npm run bootstrap:tidb
  • npm run tidb -- save-evidence-batch
  • npm run tidb -- save-opportunity-batch
  • npm run tidb -- get-opportunity
  • npm run tidb -- save-prd
  • npm run tidb -- get-prd
  • npm run tidb -- list-runs

Compatibility-only commands:

  • npm run tidb -- save-evidence
  • npm run tidb -- save-opportunity

Use the single-row save commands only for manual recovery or compatibility flows.

Normal Execution Defaults

  • Check required environment once per run.
  • Collect evidence first, then persist in batch.
  • Prefer JSON via stdin over temp files.
  • Use --input-file only when the payload is too large or awkward for stdin.

Discovery Mode

  1. Read references/contracts.md once.
  2. Check required environment.
  3. Collect evidence with search/scraping tools or skills (e.g. bundled websearch, meirkad/bright-data, buksan1950/reddit-readonly).
  4. Persist all evidence with save-evidence-batch.
  5. Cluster and rank opportunities using the contract rubric.
  6. Persist ranked opportunities with save-opportunity-batch.
  7. Return a ranked summary with run_id, opportunity_id, score, confidence, and caveats.

PRD Mode

  1. Resolve the target opportunity by opportunity_id, or by run_id plus rank when explicitly provided.
  2. Read the persisted opportunity with get-opportunity.
  3. Generate the PRD from the persisted opportunity and its evidence, not chat memory.
  4. Persist the PRD with save-prd.
  5. Return the human-readable PRD and include prd_id.

Non-Drift Rules

  • Never invent SQL. Use the documented helper commands.
  • Never generate a PRD from an unpersisted opportunity summary.
  • Never use conversational numbering like "opportunity #2" as the only identifier once results are persisted.
  • Never claim product-market fit.
  • When evidence is weak, say insufficient evidence.
  • Preserve contradictions when sources disagree.

Quality Bar

  • Prefer repeated, concrete complaints over clever one-offs.
  • Distinguish category pain from vendor-specific complaints.
  • Keep evidence traceable through source_url, pain_cluster_id, and persisted identifiers.
  • Optimize PRDs for a coding agent with explicit scope, non-goals, constraints, and acceptance criteria.