←
Roei Shoustak554 posts
Roei Shoustak avatar

Roei Shoustak

@roeishous

building @gethrbr. into AI, startups, football, fitness & history currently losing a lifelong battle with back pain

◷ Joined April 2025▧ 45 Photos and videos
Roei Shoustak avatar
Roei Shoustak@roeishous·2h

I still have a PS4 and GTA 6 spent 13 years in development just to end up running on that microwave

Roei Shoustak avatar
Roei Shoustak@roeishous·5h

Everyone is building a software factory now. Agents in. Pull requests out. The first thing every factory needs is memory. So teams build one. mem0, a vector DB, a big CLAUDE.md. That part takes a weekend. Then the real questions show up: - Who edits it? - Who is responsible when a rule is wrong? - Who checks it actually makes the agents better? - Which Slack thread changed this rule? - Two rules disagree. Which one wins? - Why does Claude Code know it and Cursor doesn't? - What stops an agent from running the command we banned? None of that is storage. It's upkeep. Every day, for every agent, on every repo. A software factory is a good idea. But a factory builds the product, not its own machines. Pick what you build.

Roei Shoustak avatar
Roei Shoustak@roeishous·14h

WILD human-md.ai what a great job

In the era of AI, what becomes the actual human / designer's job? I spent the last 2 weekends bringing my own thoughts around this question to life in a wild and hilarious scrolling experience packed with silly vignettes, eclectic characters, and surreal visual metaphors. All in, it's about 200 scroll-driven animated moments, 86 stop-motion performances, and 944 poses across 10+ chapters. It is hands-down my most ambitious animated scrolling work yet! and it took just 2 weekends. try it here -> human-md.ai

1:00
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 17

Harbor makes your agents better.

You build your business. Harbor makes your agents better.

0:34
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 17

2024: every tool I used had a dashboard. 2026: I use most of them through Claude Code. I think a lot of SaaS products will eventually be used through AI agents instead of UIs. And many tools have already opened up that possibility. Here's what I actually have plugged into Claude Code right now: Build - GitHub (gh): PRs, CI, and reviews from the terminal. - Harbor: Memory and rules across AI coding agents. - Harbor PR Review: Bugs, code quality + rule checks - Security audit: Vulnerability review. - REA: Reverse engineering apps and binaries. - Claude in Chrome: Browser automation in my own Chrome. Work - Notion: Plans, docs, content calendar. - Gmail: Drafts only, never sends on its own. - Google Calendar: Scheduling. - Google Drive: Files and docs. - Claude Docs: Living documentation. Grow - PostHog: Product analytics in plain English. - Apollo.io: Leads and enrichment. - claude-seo: SEO audits. - Easel: Charts right in the session. The funny thing is I still pay for most of these tools. I just barely open their dashboards anymore. Any tools I'm missing? 👇

Post attachment
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 17

claude writes and reviews ~90% of claude code's own code. best rule in their post: a change can never edit the checks that judge it. review rules come from the base branch, not the PR. the student doesn't grade their own exam.

Claude writes better code when it can verify its work. Checks that used to be expensive, like tests, benchmarks, and end-to-end runs, are now cheap. Here’s how the Claude Code team thinks about verification: claude.dev/blog/how-the-clau…

Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 16

building an AI startup that handles context management, guardrails, and code reviews is cool and all but my biggest flex is that i cut my own hair

Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 16

Asked Claude Code to analyze my last 5K messages. Apparently I've said "fuck" 187 times, "shit" 138 times, and "thank you" once. We've been through a lot together.

Post attachment
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 16

Agent Memory: Expectation vs Reality Expectation: "Let's add a CLAUDE.md. Reality: a pandora's box of 20+ other components opens: - Rules files per tool - Folder-scoped rules - Rules with the why - Rule ownership - Capturing corrections - Deduping rules - Conflicting rules - Stale rules - Revoking wrong rules - Recall per task - Ranking what's relevant - Context budget - Cross-tool sync - Team sharing - Private vs shared facts - PII masking - Guarded actions - Session handoffs - PR review against rules - Onboarding new devs - Measuring which rules fire - Audit trail AI made writing code easier, which made knowing what NOT to do even more valuable. But one file ≠ memory Of course, you can start with one CLAUDE.md, but these layers come up the moment you ask: - How does my teammate's agent learn what mine learned? - Which of these 80 rules still matter? - Why did the agent ignore the rule on line 140? - How do I stop the agent before it runs a migration on prod? There is a systematic way to build all of this out gethrbr.com/blog I put tons of useful data for you in here!

Post attachment
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 16

codex capacity errors went from 6.82% of daily users on the worst october day to 0.085% by nov 12 (tibo's chart). the fix wasn't a new model. it was retry delays and websocket plumbing. reliability is the feature nobody demos

Codex team might have actually fixed the "model at capacity" errors that have been pestering you.

Post attachment
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 15

Your agents might not be reading your AGENTS.md/CLAUDE.md at all Claude Code skips AGENTS.md if you have a CLAUDE.md. Codex skips CLAUDE.md unless you set it up. Neither tells you. I built a free tool that checks yours in 5 seconds: gethrbr.com/check Your agent is only as good as the file it reads!

Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 15

New Feature alert 🚨

End your session with "handoff". Tomorrow any agent picks up where you stopped. Out sick? Hand it to a teammate instead, they get it in their next session. Harbor Handoffs. Works in Claude Code, Codex, Cursor and Copilot. gethrbr.com

0:53
Roei Shoustak avatar
Roei Shoustak@roeishous·Nov 14

new coding agent index: fable 5.2 and opus 5.5 tied at 58 fable on max effort scored 56, lower than on high, because it kept hitting time limits so the expensive setting can be slower, pricier and worse what effort level do you actually leave your agent on day to day?