fangdididi

Cold Read Spec — Testing skill for Claude Code

Testing community

Write the spec first, then prove a zero-context reader can build from it: numbered requirements with worked examples, a cold-read quiz scored MATCH/MISREAD/GAP, and a spec-to-test trace that catches d.

How to install Cold Read Spec

This entry records only its repository, not the path inside it, so there is no exact command to give. Open fangdididi/cold-read-spec and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Cold Read Spec does

Alternatives in Testing

L Debug Trace — Analytically debug issues by tracing data flow through code - no logs, no test-and-see 3k ★
Test Notebook — Test the notebook currently being worked on using the notebook-runner subagent and report all results back 56 ★
Backprop — Trace a bug back to a spec gap and generate regression tests 55 ★

README

Cold-Read Spec

**English** | [简体中文](README.zh-CN.md)

An agent skill for writing a feature spec **before** coding and proving it can be built by someone who was not in the conversation.

The author of a spec cannot see its holes, because their memory fills them. So this skill has a second reader with zero context, a fresh agent session or subagent given only the spec file, answer a quiz the author wrote in advance. Every answer that comes back wrong or `NOT IN SPEC` points at a sentence to fix. Code starts only when the cold reader passes, it is written test-first against numbered examples, and before merge every requirement is traced to a test.

Why

AI agents implement exactly what is written, and guess the rest with confidence.
New sessions, teammates and future you do not share the conversation that produced the spec.
Reviewing your own spec feels thorough and finds almost nothing.
Code that quietly grows past its spec leaves nothing to check the next change against.

Features

A spec shape that can be checked: problem with a number, a "done means" check, out-of-scope fence, glossary, requirements R1.. with MUST/SHOULD/MAY, worked examples R1.a.. in given/when/then form, decisions D1.., open questions Q1.. with owners.
Lint for form: missing examples, duplicate ids, vague words such as "fast", "intuitive", "快速" or "友好" without a number, examples with no checkable outcome, unowned questions, "ready" while questions remain.
Cold-read quiz: answer key written first, a fresh reader through one of four lenses (Builder, Tester, Newcomer, Skeptic), automatic scoring into MATCH / MISREAD / GAP and the list of spec ids to fix.
Fix loop rules: fix the spec, never coach the reader; use a new reader each round; split the feature if round three still fails.
Test-first build against example ids, then a trace that flags untested requirements and DRIFT (tests naming ids the spec never had).
English a