English | 中文说明见下
find-unknowns is an Agent Skill (SKILL.md) that gives your AI coding agent a repeatable SOP to systematically discover unknown unknowns — the gaps between the map (everything you've told the agent) and the territory (where the real work happens: your codebase, constraints, and users).
Instead of letting the agent guess what you want at every blind spot, this skill forces it to surface those blind spots before, during, and after implementation — so you make the decisions that matter and the agent handles the rest.
Every agent mistake traces back to an unknown: requirements it guessed, edge cases it didn't foresee, "obvious" context that was never written down. This skill categorizes unknowns into four types and maps each to a concrete technique:
| Unknown type | Meaning | Technique |
|---|---|---|
| Known knowns | Already in the prompt | (already on the map) |
| Known unknowns | You know you haven't decided | ③ Interview |
| Unknown knowns | Too obvious to write down | ② Brainstorm & prototype |
| Unknown unknowns | Never even considered | ① Blind spot scan |
Plus two universal techniques: ④ reference material (point at source code instead of describing) and ⑤ implementation plan (final gate before coding).
- Phase 1 — Before implementation: ① Blind spot scan (dig out unknown unknowns) · ② Brainstorm & prototype (2–4 wildly different directions, fake data, no wiring) · ③ Interview (one question at a time, architecture-changing answers first) · ④ Reference material (source > screenshot > text) · ⑤ Implementation plan (decisions-you'll-tweak on top, mechanical work buried)
- Phase 2 — During implementation: ⑥ Implementation notes (
implementation-notes.md: Deviations / Open Questions / Decisions — conservative picks, keep running, only stop for architecture-level deviations) - Phase 3 — After implementation: ⑦ Pitch document (demo first, pre-answer expert objections, honest deviations) · ⑧ Quiz (you must pass a quiz on the changes before merge)
Includes a HARD-GATE: no implementation code until the plan is approved — and a red-flag table of the exact thoughts that mean you're skipping the SOP.
npx skills add ChanningYul/find-unknownsCopy SKILL.md into your agent's skills directory:
git clone https://github.com/ChanningYul/find-unknowns.git
mkdir -p ~/.claude/skills/find-unknowns # or ~/.dsh/skills/, ~/.codex/skills/, .agent/skills/ …
cp find-unknowns/SKILL.md ~/.claude/skills/find-unknowns/SKILL.mdThen restart your agent session (or start a new one) so it rescans its skills directory. Works with any agent runtime that loads Anthropic-format SKILL.md files: Claude Code, DeepSeek Harness (DSH), OpenCode, Codex, Cursor, and most agent frameworks.
Paste this into your agent:
Install the skill from https://github.com/ChanningYul/find-unknowns into your skills directory (~/.claude/skills/find-unknowns/SKILL.md), then confirm it's loaded.
Your agent will pick this skill up automatically when you say things like:
- “做个盲点扫描” / “do a blind spot pass”
- “我不知道该问什么” / “I don't know what to ask”
- “采访我” / “interview me”
- “先做原型” / “出几个方向让我挑” / “mock it first, I want to react”
- “写实施计划” / “write an implementation plan”
- “记 implementation notes” / “keep implementation notes”
- “打包成 pitch 文档” / “write a pitch doc”
- “考考我” / “出题测验” / “quiz me on the changes”
Or invoke it directly: /find-unknowns.
找出你的未知：一套围绕「四类未知」的 agent 协作 SOP，让 AI 编程 agent 在实施前（盲点扫描 / 头脑风暴原型 / 访谈 / 参考资料 / 实施计划）、实施中（实施笔记）、实施后（pitch 文档 / 测验）系统性挖出并清除「地图与疆域」之间的落差——地图是你喂给 agent 的一切，疆域是工作真正发生的地方。
- 适用场景：开新功能、进新模块、做不熟悉的领域
- 形式：单文件
SKILL.md，零依赖，装进任意支持 Agent Skills 的运行时即可 - 安装：
npx skills add ChanningYul/find-unknowns，或手动复制到~/.claude/skills/find-unknowns/
MIT