Henry (@thehqw) 's Twitter Profile
Henry

@thehqw

Another boring account with fake followers. Views expressed are my own though I don’t tweet much. OSS advocate. CTO @rss3_ & CoF @webisopen

ID: 195429894

linkhttp://henry.wang calendar_today26-09-2010 17:49:26

679 Tweet

8,8K Takipçi

296 Takip Edilen

Henry (@thehqw) 's Twitter Profile Photo

#AIxStock Did my AI make an accurate prediction on stock earning move? A yes means a correct prediction on: 1. whether it meets/beats/misses rev 2. whether it meets/beats/misses eps 3. the immediate movement (up/down) Result: Broadcom AVGO: yes, yes, no

Henry (@thehqw) 's Twitter Profile Photo

so @google gemini 3 pro is really good at vibe coding, but extremely bad at software engineering. this is an API endpoint (using fastapi) written by gemini 3 pro vs OpenAI gpt 5.2, same prompt. from my experience, gemini 3 pro lacks the "continuity" and ability to understand

so @google gemini 3 pro is really good at vibe coding, but extremely bad at software engineering.
this is an API endpoint (using fastapi) written by gemini 3 pro vs <a href="/OpenAI/">OpenAI</a> gpt 5.2, same prompt.
from my experience, gemini 3 pro lacks the "continuity" and ability to understand
Henry (@thehqw) 's Twitter Profile Photo

code must be human readable and maintainable, machine execution is not the priority. with the amount of code being generated, it should ensure that humans can review efficiently. my productivity was boosted >10x with this principle (and an AGENTS.md tailored to

code must be human readable and maintainable, machine execution is not the priority. with the amount of code being generated, it should ensure that humans can review efficiently.

my productivity was boosted &gt;10x with this principle (and an AGENTS.md tailored to
Henry (@thehqw) 's Twitter Profile Photo

have been tuning my AGENTS md (with great examples from many devs on X, I lost track of their names but a huge thank you), I think it is getting to a pretty usable state, my gpt-5.2-codex is consistently producing zero/one-shot solutions even for very large refactor work.

Henry (@thehqw) 's Twitter Profile Photo

i think GitHub copilot is very underrated, i found it extremely useful for code reviews, given the amount of code we produce using OpenAI codex and Anthropic claude code, it is pretty good at playing the last quality gate. and copilot gets RL from review outcomes, which

i think <a href="/github/">GitHub</a> copilot is very underrated, i found it extremely useful for code reviews, given the amount of code we produce using <a href="/OpenAI/">OpenAI</a> codex and <a href="/AnthropicAI/">Anthropic</a> claude code, it is pretty good at playing the last quality gate. and copilot gets RL from review outcomes, which
Henry (@thehqw) 's Twitter Profile Photo

if your code is "AI-readable only", it's a LIABILITY. mass adoption of coding agents is cutting the junior dev pipeline, if we don’t write code that humans can learn from, we’re trading short-term speed for long-term fragility. I have this sentence in my

Henry (@thehqw) 's Twitter Profile Photo

trying to use Google AI gemini with gcp Google Cloud and boy wasn't it tricky. 1.) most google searches land on ai.google.dev for Gemini API, which NEVER mentioned vertex ai at all. 2.) vertex ai studio !== google ai studio. 3.) SDK is called google-genai but env

trying to use <a href="/GoogleAI/">Google AI</a> gemini with gcp <a href="/googlecloud/">Google Cloud</a>  and boy wasn't it tricky.

1.) most google searches land on ai.google.dev for Gemini API, which NEVER mentioned vertex ai at all.
2.) vertex ai studio !== google ai studio.
3.) SDK is called google-genai but env
Henry (@thehqw) 's Twitter Profile Photo

fewer clicks for me when my AGENTS.md has this line: - Always assign reviewer to copilot via `gh pr edit [PR number] --add-reviewer "copilot-pull-request-reviewer[bot]"` thanks to GitHub Copilot i have more time for dave the diver