Find Unknowns — AI skill for Claude Code

find-unknowns — Agent Skill: Surface Your Unknowns Before You Ship

**English** | [中文说明见下](#中文说明)

`find-unknowns` is an [Agent Skill](https://agentskills.io) (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.

Why

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).

What's inside — 8 techniques across 3 phases

• 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 d