Zhanwei Wang

We Built Ourselves an AI Coding Workbench — Using GitHub Actions as the Sandbox

Fri, 08 May 2026 00:00:00 +0800

How it started

Over the last year or two, our team has been doing more and more coding work with AI. We started by running CLIs like Claude Code locally, and quickly hit a practical problem: every agent task I kicked off held my laptop hostage. The agent was building, running tests, editing files, thinking about the next step; meanwhile I wanted to do something else, and the CPU fan started screaming. Forget about running five agents on five unrelated things at once.

"All Tests Pass" Doesn't Mean "Requirements Are Met" — Why AI-Assisted Development Is Often Suspiciously Green

Mon, 27 Apr 2026 00:00:00 +0800

Many teams that delegate long-running tasks to AI collaborators keep running into the same uncanny ending:

The task wraps up. CI is fully green.
The test report shows hundreds of new specs, beautiful coverage, smooth flow.
A few days later, the product ships, and the first real user walks through a core flow — only to trigger a bug that should have been covered but wasn’t.

You go back to the code and find: that “covered” test was actually just a status-code probe; that “edge case handler” was a console.warn followed by return; that “complete feature” is missing a database table, a cookie, a middleware — but none of those gaps make any test fail.

Cost and Performance Optimization for Claude Code Skills: 6 Principles from a Real Session

Fri, 17 Apr 2026 00:00:00 +0800

This article is based on a week-long optimization effort across three production Skills — prd-analysis, system-design, and autoforge — from token-level measurement to actual code changes. All numbers come from real JSONL session files, with inflation factors corrected.

Why skill cost deserves its own treatment

Generic “LLM cost reduction” articles usually talk about context pruning, cache warmup, and model downgrading. These apply to Skills too — but the Skill execution environment has several structural differences:

When Documentation Stops Rotting: From Karpathy's LLM Wiki to Software Engineering Documentation

Thu, 09 Apr 2026 00:00:00 +0800

Humans write the original documents and make decisions. LLMs handle synthesis, updates, and consistency checks. Documentation is no longer a static artifact that rots the moment it’s written — it becomes a living knowledge base continuously maintained by LLMs.

An age-old problem

Every software engineer has experienced this: you join a project, open the docs, and find the README still describing a module deleted six months ago, API docs with fields that don’t match the code at all, and architecture diagrams showing a design from two versions back. You ask a colleague, who says, “Don’t bother with the docs — just read the code.”

Frontend-Driven Development in the Vibe Coding Era

Mon, 09 Mar 2026 10:00:00 +0800

When AI becomes a routine tool in development, how should we change the way we work?

The evolution of three development models

Before discussing frontend-driven development, here are three different development models and how they’ve evolved:

1. Backend-driven development (the traditional way)

Product requirements → Backend designs API (1-2 weeks) → Frontend builds pages (2 weeks) → Integration

A real pain point:

Last year I worked on an e-commerce admin system where the team followed this model. By the time the frontend received the API documentation, we were already in week three of the development cycle. I spent two weeks implementing the order management module according to the docs. When we demoed it to the product manager, she frowned: “Why can’t the order list be filtered by status? Users definitely need this.”