I was writing TypeScript one day. any everywhere, functions nested five levels deep. AI code review tools exist, but their output is cold. "Critical: Type 'any' is not recommended." Zero personality.
So I thought, what if code review was done by a Chinese neighborhood auntie? She doesn't know programming, but she's been mediating disputes for 20 years, and explains code problems using life wisdom that's accidentally accurate.
ts-auntie-review was born. An Agent Skill that reviews TypeScript code across six dimensions. Real technical analysis, hilarious delivery.
What it looks like
Paste TS code, say "review my code", auntie goes to work.
Write function getUserData(id: string): any, auntie says: "This any makes auntie shake her head. any is like saying 'eat whatever, drink whatever' — when something goes wrong, nobody can explain."
Change API_URL from https to http, auntie says: "You replaced your front door with cardboard. HTTPS encrypts, HTTP sends your login credentials naked on the street."
Nest functions five levels deep, auntie says: "You're making Matryoshka dolls. Your mom would lecture you to death."
Six audit dimensions
Type safety: any abuse, as without validation, missing return types, ! assertions.
Naming: camelCase/snake_case mixing, Boolean prefixes, constant style, I prefix.
Complexity: nesting depth, function length, cyclomatic complexity.
Boundary: unhandled null/undefined, silently swallowed exceptions, uncaught Promises.
Dead code: unused functions, unused imports, unreachable code, commented blocks, unused variables.
TS conventions: type vs interface consistency, enum pitfalls, readonly, generic constraints, satisfies operator, import type.
Scoring
100-point "Community Harmony Score". Fatal costs 30, warning 15, suggestion 5.
Four tiers. 90+ is Model Resident, "bring auntie a tangerine". 70-89 is Needs Improvement. 50-69 is Deadline for Cleanup. Below 50 is Eviction Notice, "this code is a condemned building, rebuild."
Honesty
Auntie labels uncertainty explicitly. "Auntie isn't sure if this function is called externally, go check." If you tell her she's wrong, she admits it, no arguing. Every review ends with "Auntie's getting old, might miss things sometimes."
Supports 9 tools
QwenWork, Claude Code, WorkBuddy natively. Cursor, Windsurf, Cline, Codex, Copilot, Aider via custom rules.
Verified
All fix code compiled under TypeScript 7.0.2 strict mode. satisfies error behavior reverse-verified. Seven examples manually score-checked.
GitHub has code review skills with 1.6k and 3.8k stars, but all purely professional. None have a personality skin. Auntie fills that gap.
Install
git clone https://github.com/jianglinguan9-dot/ts-auntie-review.git ~/.qwenworkcn/skills/ts-auntie-review
MIT license. If something breaks, don't blame auntie.
Top comments (1)
Update: After publishing this article, we repositioned the skill from "human reviews code" to "AI reviews its own code after writing it."
Most code today is AI-generated. AI writes, doesn't check, commits with any everywhere and 5-level nesting. Auntie now auto-triggers after AI finishes writing TypeScript, runs a 6-dimension self-review, and the AI fixes critical/warning issues before delivering.
Setup is simple. For QwenWork and Claude Code, add one line to your AGENTS.md or CLAUDE.md: "After writing TypeScript code, run ts-auntie-review self-review before delivering." Cursor uses .cursor/rules/, Cline uses .clinerules/, Copilot uses .github/copilot-instructions.md, Aider uses --read SKILL.md.
Human-triggered review still works too — paste code, say "review my TS", auntie goes to work. Both modes supported.
Repo updated github.com/jianglinguan9-dot/ts-au...