Backend in a box
plus vibe-coders are taking over at real companies
The newsletter for ai builders of all levels. Mini-tutorials, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks,
ChatGPT finally rolled out ‘project-only’ memory. ie, it’ll only use memories its remembered from the project - and not bring in general memories from your total ChatGPT profile.
Claude Code has a new /context command to visualise your context window and token usage.
Some people are turning their back on Claude Code for Codex CLI (by OpenAI) - it’s been neglected as a product recently but folks are getting great results with GPT-5. And it sounds like a lot of work is going into the product experience.
Agentic browsers like Perplexity’s Comet remain vulnerable to indirect prompt injection attacks. Which is why folks like Browserbase push for “1. your agent should have it's own browser with limited access to it's information, and 2. it should run far away from your data.” - this thread highlights more of why it’s a massive issue.
I backed a company recently that’s in the current YC batch. The reason? Because they strip all the pain of auth, database, email, observability etc with an all-in-one Typescript framework. It’s like a backend in a box (which is something I’ve been posting about a lot). Check Modelence out and give it a spin for your projects.
Is your enterprise data at risk from unchecked AI use? Airia lets you rapidly deploy AI agents at scale across your enterprise - all with built-in governance, runtime security, data leakage prevention, and more. Plans start at $49/month. Secure AI at scale.*
*sponsored
Last minute ad slot available. Feature your tool in this Thursday’s post for 40% off.
🌐 What I’m consuming
A journalist spent two days vibe coding at Notion and also shipped some actual code.
The context window problem: scaling agents beyond token limits.
Has AI gotten good enough to predict my taste?
YC’s collection of talks on Context Engineering.
Building and prototyping with Claude Code’s PM, Cat Wu.
We put a coding agent in a while loop, and it shipped 6 repos overnight.
The full journey of vibe-coding something real in three weeks.
How to get your pitch for candidates right in the AI talent wars.
Practical example of refactoring an existing codebase with AI.
⚙️ Tools to tinker with
Triple Whale's Moby Agents - Drive smarter, faster decisions across your entire business with the only AI agents built for ecommerce.*
Disco - Super customisable AI integration that lets you toggle tools on/off. Free in preview.
Synclabs - generates studio-grade lip sync in minutes with their new model lipsync-2-pro.
Claudable - Open source web builder that uses CLI agents to build products.
Humanlayer - Build agents that feel like real coworkers.
Rube by Composio - Connects your AI tools to 500+ apps instantly.
Describe Audience by SparkToro - Describe your audience in plain English, and get a report on that audience's behaviours and demographics.
Montra - A slide based video creation tool that makes it easy to create professional videos.
*sponsored
🥣 Dev dish
VibeVoice by Microsoft: A frontier long conversational text-to-speech model.
CCPM - Claude Code workflow to ship better using spec-driven development.
uwu - natural language to shell commands. like making your own ‘Warp’ terminal in any terminal!
Vercel AI Gateway is now generally available. Zero markup on tokens, including BYOK.
Fractal - A desktop web browser that generates website content using LLMs (no fetching from the internet).
poem-cam - Capture photos and generate cute, cursive poems for each shot.
colf.dev - Prompt your way to the solution with the fewest tokens used, given a set of code challenges.
🍦 Afters
Anthropic’s higher education advisory board. + Claude for education (podcast).
OpenAI’s new CEO of Applications frees up Sam Altman to focus on GPUs, brain-computer interfaces, and consumer hardware.
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumbnail creds: @keshavatearth