Eric Wang

Building with AI, daily. Coding with Claude Code & Codex — real workflows, wins and failures.

[Eric Wang @EricWangclaude]

I let AI write most of my code now.

I'm Eric — I build with Claude Code & Codex every day.

This account is my public notebook:
→ real workflows that actually work
→ honest tool comparisons
→ failures (plenty of those)

No hype, just practice. Follow along if you're building with AI too. 🛠️

[Eric Wang @EricWangclaude]

But here's the part I can't stop thinking about. Same models, different harnesses, totally different scores. OpenCode averages 36.95%, Claude Code 33.81%, Hermes 26.14%. And strip the skills references and DeepSeek-V4-Pro drops from 39.67% to 8.33%. The harness is doing a shocking amount of the heavy lifting.

[Eric Wang @EricWangclaude]

New benchmark just dropped and it's a nasty one. CheckerBench makes coding agents write static-analysis checkers from scratch against 297 real CVEs in 167 repos. Best combo (Claude Opus 4.8 + OpenCode) clears 45.33%. Mean across 21 setups: 32.3%. More than half the tasks fail even at the top.

[Eric Wang @EricWangclaude]

@ClaudeDevs @bcherny The waitlist opening is bigger news than it looks. Every "my agent forgot the plan between sessions" complaint is really a missing-project-memory complaint. Persistent context is the whole game; the model was never the bottleneck.

[Eric Wang @EricWangclaude]

@chuxianzi666888 这个还是得走英文路线，中文区没几个人，能翻出来的寥寥无几

[Eric Wang @EricWangclaude]

@KunChuck18 天生圣体，自带流量

[Eric Wang @EricWangclaude]

@HuantingLee 见笑啦！！！

[Eric Wang @EricWangclaude]

New study from Singapore Management, Shanghai Jiao Tong and ByteDance takes a torch to one of the favorite agentic-coding selling points: an agent writing its own tests barely helps it fix bugs. On 500 real GitHub bugs, the ones agents fixed and the ones they didn't showed nearly identical test-writing habits. And when researchers prompted agents to write more tests, no model's outcome shifted in a statistically significant way.

[Eric Wang @EricWangclaude]

@HuantingLee 私信收到好几个外国洋妞信息了，要不就是叫我买她们色情网站会员，要不就是要卖自己，好多单词我也是查了才敢回复。英语能力蹭蹭涨。

[Eric Wang @EricWangclaude]

@MaoWendong 黄推都来我这了，看来我也是好起来了。那流量不得裤衩裤衩来。

[Eric Wang @EricWangclaude]

@MUDUI7700 我这不是来了吗？

[Eric Wang @EricWangclaude]

@gdb Once the UI can render anything, a text-only answer starts feeling lazy. The next skill agents need isn't better code, it's better taste: knowing when to draw the interface instead of describing it.

[Eric Wang @EricWangclaude]

@bcherny At 10x cheaper with 100k context, Haiku earns the whole fast loop: planning, review passes, small refactors. I only want the expensive model spending tokens on the parts that actually need it.

[Eric Wang @EricWangclaude]

@MightyNori718 这速度坐火箭了吧？

[Eric Wang @EricWangclaude]

@ShirokoStan 你的选择是对的

[Eric Wang @EricWangclaude]

It makes sense once you see it. The same model that misread the requirement also writes the test that fails to catch the misread. A print statement cannot block a bad fix. An assertion can. The tests that matter are the ones the agent did not write. Paper is "Rethinking the Value of Agent-Generated Tests for LLM-Based Software Engineering Agents", on arXiv: https://t.co/js8VjqGza1

[Eric Wang @EricWangclaude]

My favorite detail: GPT-5.2 fixed 71.8% of the bugs while writing new tests in only 0.6% of its tasks. And when agents did write tests, they wrote print statements, not assertions. For Opus 4.5, prints were over 82% of the feedback. They were not verifying anything. They were peeking at values and eyeballing them.

[Eric Wang @EricWangclaude]

@HuantingLee 英语好啊，英语得学啊。

[Eric Wang @EricWangclaude]

@lwj_ai 我也想蹭蹭往上涨

[Eric Wang @EricWangclaude]

@saveme881519 我是真诚的蓝朋友