Shivam @shivambuild

anthropic.com/engineering/re… — Anthropic's own account. The truck is a metaphor; it had already tried partial Git features and mirrors. Performance numbers are self-reported.

25 million code checkouts a day. One online CI job asking for a file. Anthropic moved history work to a server and fetched files on demand; it reports checkouts dropping from minutes to seconds. Git still fits most repos.

Product workflow: spryloom.com/docs/publishing . The tracker is our internal app; the characters and posting mark are dramatized.

"I thought you posted it." "I thought you posted it." We built a video tracker with an agent, published it with Spryloom, and invited our team privately. Email link in, mark each channel, done. The status tap in the skit is a demo. What would your team build?

Source: Anthropic, "Project Swap" (Sep 24, 2026). Employees only, cooperative Claude agents, one market for books. anthropic.com/research/proje…

Anthropic sent 201 AI agents to trade books for their people. They negotiated fine. What hurt was understanding the people: in this experiment, 85% of the shortfall from the best outcome came from misread preferences, 15% from the trading.

Who should be able to open the app your agent built? Your agent finished the app. Before you send a link, decide whether it is for you, a few named people, or your whole company. Here is what each choice lets people do and how to change it later. spryloom.com/blog/who-shoul…

Client four shouldn't be harder than client one. Tag the consultant who just got "can you resend the link?" again. #claudecode #aiconsulting #agencyowner

One rule for this experiment: if the app needs a training session, I've probably made it too complicated. The goal is a tool a team can understand in 30 seconds. What would you want on that first screen?

Imagine opening one page on Monday and seeing: • 3 things waiting for your approval • 2 tasks blocked by someone else • 1 deadline slipping No digging through email or asking for updates. That's the kind of focused app I want to build with Spryloom.

I keep noticing the same thing in small teams: Nobody wants another project management tool. They just want to know: • What's waiting on me? • Who owns the next step? • What's stuck? This week, I'm exploring a tiny team tracker that answers just those 3 questions.

Our first real Spryloom user isn't a developer. And that might be the most exciting thing that's happened since we started building. AI is changing who can create software. We're betting the next wave of apps won't come from software companies.

The cost of building an MVP keeps falling. The cost of getting someone to care about it hasn't. Feels like distribution is becoming the real technical challenge for founders.

People using Claude Code daily: What's something you built recently that you genuinely wouldn't have attempted a year ago? Not a demo. Something you actually use.

@zahragr8r You should be able to publish your things to spryloom.com

Unpopular opinion: We're going to see a lot more bad software before we see a lot more great software. AI removed the barrier to building. It didn't automatically give everyone good product judgment.

Claude Code has changed my expectations. Waiting two weeks for a simple internal tool suddenly feels ridiculous. I wonder how many company processes exist only because software used to be expensive.

Everyone is celebrating how fast AI can write code. I'm more interested in what happens when writing code is no longer the expensive part. What becomes the new bottleneck?

Something I'm learning while building @spryloom: AI has made it easier than ever to build a product. It hasn't made it any easier to get your first 10 users. The bottleneck keeps moving.

The surprising part wasn't the numbers. It was watching the agent stop for things only a human could do: sign up, paste a key, click through a Google consent screen. That last mile is the slow part now. We're building Spryloom to fix it. x.com/spryloom/statu…

spryloom @spryloom

2 weeks ago

We had Claude Code ship the same app three times. Only the route to a coworker changed. Steps a person had to do before the coworker's request was approved: Spryloom: 1 Cloudflare: 5 Vercel + Supabase: 11 Every stop is published 👇

1 0 0 18 0