How I build products, after a decade of doing it:
1. Design first. If it isn't clear, it isn't done.
2. Talk to customers every week. Rule of 100
3. Architecture is a product decision.
4. Train like an athlete: consistency beats intensity.
5. Build for the next ten years, not the next demo.
That's how I'm building @linefoldapp. If you lead product or engineering, my DMs are open.
Our CI took 10–25 minutes a run. One prompt got it under 2.
Four changes:
1. Typecheck, 3 test shards and the image build run in parallel
2. Docker builds natively per arch. No QEMU
3. The image reuses the bundle CI already built
4. All 14 test groups run at once
The prompt 👇
hill climb will need one of the most crucial inputs: culture.
It defines the timing of the world and what’s the public pulse. this one will be hard to get right in the first go since lots of AI generated content is flooded across platforms so any sort of agent generated culture metric will be false data.
just like data labelling ultimately this is where humans will be still valuable, defining their conversations with their friends, families and coworkers to the agent as an input for the hill climb.
@zachlloydtweets Internally in linefold we use:
1. Customer issues reduction % per week
2. Each PR is linked to an initiative (inside linefold) and each initiative generates a weekly report wrt to % MRR increased so features like greater user onboarding, landing page perf optimizations win 🥇
You can now run coding agents on your own computer.
Install the linefold CLI and run linefold pair. Sessions run in the folders you share, on the Claude Code, Codex or Cursor sign-in you already have. No open port, no tunnel.
How my day works now:
· 2–3 hours grilling the spec
· delegate the issues to agents in Linefold
· end of day: verify and review what shipped
Last 30 days: 323 PRs merged across 5 repos. ~11 a day. 97% of 903 commits co-authored by Claude.
The spec is the product now.
The prompt, trimmed:
"Our current CI timelines are in the order of 10-20 mins each. I want to get this down to <1-2 mins.
Give me all possible scenarios. We are fine with sharding, or even skipping checks if there are no changes.
Don't trigger any build on GitHub Actions. Run everything locally and benchmark it, replicating the GitHub Actions environment."
Our CI took 10–25 minutes a run. One prompt got it under 2.
Four changes:
1. Typecheck, 3 test shards and the image build run in parallel
2. Docker builds natively per arch. No QEMU
3. The image reuses the bundle CI already built
4. All 14 test groups run at once
The prompt 👇
this math is exactly why we shipped Agent37 as a cloud provider in Linefold last week.
bring your own key and every coding session runs in its own fresh sandbox, billed by the minute to your Agent37 wallet.
race to the bottom? we're in 🏁
The CEO of Vercel called us out. We claimed https://t.co/DqCfjvVkFz was 25x cheaper, and he said we didn't consider active CPU time.
So we fixed the math..
We gave 100 coding agents a sandbox each and had them vibe code an app. Here's the cost per hour at just 5% active CPU:
Vercel: $9.75 (14x)
E2B: $16.56 (24x)
Daytona: $16.56 (24x)
Agent37: $0.68
We're starting a race to the bottom... see you there.
Find a cheaper sandbox and we'll refund the difference.
On X, a reply is worth 7 likes.
A reply from someone you follow back is worth 28.
From the open-sourced ranking code: reply 5.0, like 0.708, mutual-follow reply boost +15.
So stop asking for likes. Ask a question people want to answer.
@dani_avila7 Cleanup is a big part of this. Your brain needs to forget old ones. It can happen only when its removed from the view.
Always archive your done sessions. @linefoldapp lets you do this easily to help your focus.
Group-chat likes don't help your post on X.
X's own comment in the ranking code: "Directly navigating to a post (i.e., coordinating via groupchat) has no ranking impact."
Engagement only teaches the algorithm when it happens on a post served in someone's feed.
Two major points I’d highlight:
- This space is very early and lots of great patterns everywhere. Pick one which resonates with your team and move ahead
- Focus on closing the loop first before jumping into quality of outputs directly. I’ve personally found that with a decent codebase obsessing over the evals wasn’t much helpful
I’m building @linefoldapp for this exact use case. Teams see what every session is doing with support for Private teams ofc. And Loops helps you bump up the shared context at the end of each day. Insights help you understand how you must tweak your AGENTS.md and also build routines. It’s magical tbh
Would be more than happy to give a demo :)
Last 30 days on Linefold:
255 PRs merged
1,886 commits
99% of them co-authored by Claude
My job moved from writing code to writing specs and reviewing PRs. Review is the bottleneck now.
What's yours?