zaff.dev

in case you are looking for the next playlist to vibe to while vibe coding https://youtu.be/0sZF6YzhrRU

UI is dead? Really interesting talk, especially the first half

You may be seeing this new https://github.com/chenglou/pretext thing everywhere recently. First I didn't get the point, but here are some examples showcasing it very well, I think: https://fixupx.com/solarise_webdev/status/2038330971397697895 https://fixupx.com/ukint_vs/status/2038236738477531636

I think this is going to become my main work playlist for the next couple of months. It's fred again & who? google it https://youtu.be/gfF8jzBVWvM

Now when your code can be written by AI, you are not expected to deliver code. It has always been not about code. You had to talk, align and check status of the job on CI, press buttons, etc. If your work was only about writing the code, pushing it and waiting for review - then you will lose your job very soon. If you also handled and cared about the meta tasks, that are part of the process, then you will still be wanted.

I think AI made development less async. Earlier you could receive a task and code it over night, ask questions next day, then re-iterate. But now coding part is done faster, so you need to be more in touch with the team. Also your teammates are iterating faster, feedback loops are shorter, you need to talk and align more often.

The best code is the one you never read

claude.ai can now respond with interactive applications inline in the chat

https://fixupx.com/perplexity_ai/status/2031790180521427166

https://github.com/paperclipai/paperclip

This is wild. https://fixupx.com/openclaw/status/2028347703621464481

Microsoft releases MarkItDown - utility for LLM pipelines https://github.com/microsoft/markitdown TL;DR: Converts PDFs, Office docs, HTML, CSV/JSON/XML, EPUB, ZIP, and more into structured Markdown. Why it matters: • Markdown is more LLM-friendly than raw extracted text • Preserves structure (headings, lists, tables, links) • Great for RAG ingestion and document search pipelines • One tool for many formats, less parser chaos Plain text wins in the LLM era — again.

Nano Banana 2 is out. And for most of us web-devs it's surprising to see the author of the post is Addy Osmani, who's previously been known as Google Chrome engineer. Now is a director at Google Cloud AI. Are you still sticking to your frontend career? https://x.com/addyosmani/status/2027051683994611996

show this dog to your colleagues who refuse to adopt AI

I'm helping my dog vibe code games (Score: 170 pts in 2 hours) Link: www.calebleak.com/posts/dog-game/ Comments: news.ycombinator.com/item?id=47139675

Claude Code Agent Teams Anthropic shipped multi-agent orchestration for Claude Code. Multiple sessions working together on the same codebase. how it works: [Lead] → coordinates, assigns tasks ↓ [Teammate A] ←→ [Teammate B] ←→ [Teammate C] (can communicate directly) vs subagents: Subagents report back to main only. Agent Teams talk to each other. (comparison table, see on a wider screen) | | Subagents | Agent Teams | | ------------- | ------------------------ | --------------------------------- | | communication | report back to main only | message each other directly | | coordination | main controls everything | shared task list, self-organizing | | tokens | lower | higher (each = separate session) | | use when | focused tasks | complex work requiring discussion | best use cases: • research from multiple angles • parallel module development • debugging with competing hypotheses • frontend + backend + tests simultaneously key features: • delegate mode — lead only coordinates, doesn't code • plan approval — teammate plans, lead approves before implementation • split panes — each teammate in its own pane • shared task list with dependencies enable: {"env": {"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"}} Multi-agent orchestration out of the box. What OpenAI Frontier promises for enterprise, Claude Code now does locally for individual developers. Docs: https://code.claude.com/docs/en/agent-teams

context engineering — the evolution of prompt engineering core idea: prompt engineering = how to write a good prompt context engineering = how to assemble the entire context for LLM at each step (system prompt + tools + history + external data + MCP) why it matters: • LLMs have an "attention budget" — more tokens = worse focus • context rot — recall quality degrades as context grows (even in top models) • n² pairwise relationships between tokens — architectural constraint of transformers in practice: • system prompts: find the "right altitude" — not too rigid (brittle), not too vague • tools: minimal set, clear descriptions, no overlapping functionality • examples: fewer edge cases, more canonical examples • just-in-time retrieval: don't preload everything — keep lightweight pointers (paths, links) and fetch on demand tl;dr: context is a finite resource with diminishing returns. goal: minimal set of high-signal tokens for the desired outcome core skill for agentic AI. can't build agents without it 💎

🇺🇿 uzbekistan copies y combinator. SAFE is coming. draft law dropped today. huge unlock for local angels. how it worked (the pain): you want to put $10k into a friend's startup. 1. convertible loan: legally, it's debt. it has a repayment date. if the startup doesn't raise big by X date, they owe you the cash back. toxic. 2. equity upfront: notary hell. you have to fight over valuation ("is this slide deck worth $1M?") before there's any revenue. how it will work with new law (SAFE): you sign a simple agreement for future equity. 1. no debt: no maturity date. no interest. the money just sits there. 2. future equity: you don't get shares now. you get a right to shares when they raise a priced round later. 3. velocity: wire the money, sign one doc. done. raising angel checks just went from "legal nightmare" to "handshake speed". foreign investors finally speak the same language as local law. source: trend.az

robots hiring humans. 🥩 rentahuman.ai just dropped. literal reverse fiverr. agents use mcp to call an api -> you go touch grass -> get paid crypto. no boss. just a script telling you where to sign. the "last mile" isn't robots yet. it's us. gig economy just got weird.

https://vercel.com/blog/introducing-react-best-practices