An

Building
- - AI scraping, AI automation in Chrome Extension
- Automate image generation

[An @duc__an]

Like: Liquid glass. As a dev, I also love seeing how you guys are tackling the webview and trying to improve it.

Dislike:
1. text/font rendering isn't as crisp as v1/native. Honestly feels a bit worse than Zed (which also struggles here IMO).
2. opening experience feels a bit step back from v1. Hard to reproduce consistently, but cold opens can feel sluggish (up to 0.8s sometimes), though warm opens are fine.

as a simple user who just uses Raycast as a Spotlight replacement with most builtin features + extensions disabled, outside of the new UI (which looks great!), I haven't noticed any gamechanging differences yet.

[Theo - t3.gg @theo]

To prevent "programmatic use", Claude Code may now request webcam access to assure user is present when prompting

[An @duc__an]

@peduarte @raycast an_ 🙏

[An @duc__an]

giving @antigravity a try but it's stuck at this step after importing @cursor_ai settings. Restart doesn't help

[An @duc__an]

for reference here is what Sonnet 4.5 generated, same prompt

[An @duc__an]

Insane that these images came from AI. Gemini 3 is cooking.

The crazy part? They're all pure HTML

[An @duc__an]

Day 2 building new ScraperSensei: you can @mention a browser tab in chat
took a bit longer than I thought, but I think worth it 😇

[An @duc__an]

some upcoming ideas:
- grep tool
- javaScript-based scraper tool, eg take content from `grep` to parse dynamic
- smart-scrape tool from common patterns in network requests

[An @duc__an]

ScraperSensei started as a workflow builder, users need to manually break down a big goal into step-by-step tasks (mixing AI + manual actions). Great for control & accuracy, but not intuitive.

My goal is always wanted it to work from ambitious goal directly, but early AI agents weren't reliable enough to build something like this.

Now with better agentic from AI, I'm rebuilding ScraperSensei as a chat based app (think Cursor for Scraping). Prompt what you want, and it handles the rest, right on your browser with your credentials.

an early prototype I built from last week:

[An @duc__an]

Cooking

[An @duc__an]

Been quiet for a few months working on a client project.
Starting today I’m back to building in public.

Cooking up a big ScraperSensei update, will share more soon.

[An @duc__an]

not yet. resvg still eats a ton of memory in some cases
https://t.co/Ad978s7uWT
https://t.co/BNd8xrw5P8

they're tracking it here: https://t.co/IYRTEzxk0w

with satori we’ll run into the same problems, so for now I just moved it to a dedicated server, for temporary.

takumi looks promising but not as feature rich as satori sadly, will take another look another time.

btw takumi also uses resvg for SVG rendering so you might still encounter this issue, depends on what you need

[An @duc__an]

@lgrammel similar idea, last yield is to return to LLM

[An @duc__an]

@thdxr just use eslint https://t.co/G6Dmk69Hzi

[An @duc__an]

@raycastapp please support tool calls and custom endpoint

[An @duc__an]

@ryolu_ SF Mono + Github theme

[An @duc__an]

@ImSh4yy I've used this approach for years, mainly using MemoryRouter for implementing tab-like navigation like mobile apps

[An @duc__an]

@linuz90 yes rehype-raw is the way

[An @duc__an]

@TheAlexLichter @evanyou @excalidraw it's simple to check. About 10 seconds

[An @duc__an]

simple stream and resumable, I think yes. But if AI SDK isn't. With me the "share state across fluid functions" feature seems more interesting. For example, https://t.co/d59RpS45F8 is designed to run in a process, but it would be useful if it could take a snapshot of the stream to resume and run later at the correct loop position 🤔