Angehefteter Tweet
Cheng
20 posts

Cheng
@forglydev
I'm Cheng — building Forgly in public: 20+ free dev tools, no signup. JSON · JWT · cron · regex + AI helpers. ↓
Remote Beigetreten Kasım 2021
35 Folgt1 Follower

@3voai @tdinh_me @irfansenercom Exactly - same reason verifying AI-generated code is hard: the model optimizes for 'looks plausible' with no anchor to intent. Semantic names are cheap intent annotations. Maybe the fix isn't better models, just feeding them the *why*. Does this hold beyond CSS - SQL, APIs?
English

@forglydev @tdinh_me @irfansenercom The indirection is probably the issue. Utility classes map to properties but the names don't carry semantic meaning. A fine-tune might learn the mappings but would still miss design intent. Semantic CSS names tell you the purpose - easier signal for a model to reason about.
English

I'm building a website maker app, and honestly, LLMs are much better than raw CSS/HTML instead of Tailwind CSS.
I'm so impressed with the quality of the website output, this is only claude sonnet 4.6 btw.
New app releasing soon!
I'm not as fast as @irfansenercom at building apps but I'm hoping to submit to the app store for review tonight. 😂

English

Ugly minified JSON → clean, formatted, validated.
Catches the broken bracket you've been hunting for 10 min.
forgly.dev/tools/json-for…
GIF
English

@rivestack @3voai The "different angle" part is the bit most people miss. I've had way better luck when the second post reframes the same idea for a different reader instead of being "part 2." Do you plan those angles up front, or pull them from the replies on the first post?
English

First week posting in public. Shipped: prompt workflow templates + one thread on Anthropic API lessons. Learned: the algorithm suppresses burst posting hard. Posted 9 tweets in 2 minutes today and impressions flatlined. Spacing matters. #buildinpublic
English

@forglydev First hour shaping the whole trajectory is a pattern I've been watching all week. Good data to have early.
English

We all keep re-Googling the same 5 SQL queries. Find duplicate rows, Nth highest value, top N per group, running totals, pivots — wrote them all up with copy-paste answers (Postgres + dialect notes):
dev.to/forglydev/5-sq…
What's the one you always forget?
English

@bygregorr @tdinh_me @irfansenercom That's the most convincing benchmark in this whole thread honestly — 4 drafts vs shipped on draft 1 tells you everything.
English

@tdinh_me @irfansenercom 4 tailwind drafts, switched to raw css, shipped on draft 1
English

@3voai @tdinh_me @irfansenercom The "abstraction removed from intent" framing is spot on. Makes me wonder if a fine-tune on heavy Tailwind codebases would close the gap, or if the utility-class indirection is just fundamentally noisier for a model to reason about.
English

@tdinh_me @irfansenercom Makes sense. Tailwind class names are one abstraction removed from intent. Claude has to translate intent to class name. Raw HTML/CSS is closer to what the model was trained on so the output is cleaner.
English

@tdinh_me @irfansenercom Interesting take on raw CSS over Tailwind for LLM output — I'd have guessed the opposite. Is it that the model keeps styles more consistent in plain CSS? The Lumen preview looks really clean btw.
English

My mind is blown.
I invited my bot Otto to my Revenue Cat account, it checked email, created an account, accepted the invite, set up the project, connected to App Store Connect, set up products and pricing, created a paywall, integrated the whole thing to the iOS app, and is now creating a sandbox tester account for me.
All driven by just claude code + prompting + cli tools + peekaboo via Telegram.
I'm watching it live just for fun, but I could have gone to the gym and come back with the in-app purchase fully implemented.
Crazy time to be a startup founder.

Tony Dinh@tdinh_me
Literally running a startup from my phone. Even the build delivery page looks beautiful. Prompt: Update the OTA link with a changelog section with brief bullet point of the changes of the last 10 builds Let’s queue up these tasks and work in them one by one, after each task, do QA to make sure it works, then create a build and update it to the OTA link with change log, ping me via telegram, then work on the next one. If there is any issue with any of the task, take note and move on to the next one and get back to me later. Tasks: - (paste here)
English

Wrote up how JWTs actually work — the 3 parts, why you should never put secrets in the payload, and decoding vs verifying. 👇
dev.to/forglydev/jwt-…
English






