I'm Yev. I build the tools I wanted once I let AI agents do real work for me: browse, code, clean up, make ads.
Most of them started as my own mess in the ugly middle. One reply per project below: the pain it fixes and a link.
@brocmayo Agree, the summary is lossy either way. Picking at a break just loses less than an auto summary mid-task. Decisions go in files for me too, the /compact note covers what hasn't made it there yet.
Claude Code reworded one warning this week. At 80% full, /context now says: "Auto-compact will soon summarize older messages. Run /compact now to choose what's kept."
That second sentence is the real tip. Don't let the auto summary decide for you.
https://t.co/CnoCghyGCY
@not_sacke Not guaranteed. Each compact is a fresh summary, so a second one can drop what the first kept. I repeat the keep list every time, and anything that must last goes into a file like CLAUDE.md.
ChatGPT just dropped the 10-minute dictation limit. Live Dictation shows your words as you talk, on every plan.
Talk through a messy idea for 15 minutes, then ask it to sort it out.
What would you dictate first? #ChatGPT#AInews
My habit: at a break, /compact and name what summaries drop. "keep the exact error, file paths, the approach we rejected."
Or fast-jev-compaction, my fork of Tamara Tran's plugin: every call stays, stale output becomes a note.
https://t.co/XYujrc1iyf
What do you lose?
My 3 checks: can I read every note as a plain file, delete one by hand, and copy the lot to another machine?
turbo_quant_memory passes all three: typed notes on my own disk, for Claude Code, Codex, Gemini CLI or Cursor. Free, MIT. https://t.co/bJgR9wtxPN
Where does yours live?
AWS made agent memory a managed cloud service yesterday: Amazon OpenSearch Memory is now GA.
Handy. But before I trust any memory store, I ask one thing: if that account gets suspended on Monday, what does my agent still remember?
https://t.co/R79LgYosSu
Their rule, now mine: no dismissal counts without the file and line that proves it. And don't ask the same model, it mostly agrees with itself.
consilium asks a different AI in one command, never edits my files. https://t.co/vhar2hvpzQ
Worst "handled elsewhere" you've seen?
The worst bug in AI code review isn't the one it misses. It's the one it finds, then talks itself out of.
Google's Sashiko, which reviews ~1200 Linux kernel patches a day, calls these "shaky self-dismissals": assuming an unseen caller or lock saves you.
https://t.co/BOno1L2KY6
I'll try the local one for public pages. What I actually need scraped sits behind logins: supplier portals, ad dashboards, my own SaaS admin. For that I run the agent in Murable Browser. It lives inside the browser and uses my signed-in profile, no cookie exports. https://t.co/1a0C5A6PoL
Forgelix picks how it shows up: a quiet chat, an on-page prompt or ready questions. Agents review every dialogue for mistakes, and you can take over any chat.
Try it free for 30 days and see the results for yourself. https://t.co/KOFz5lLvug
What makes you close a site chat?
Best line in a shopping AI launch this week: when nothing fits your taste, it stays quiet.
That's Croissant's new Remy agent. Most store chat widgets do the opposite and pop up on every page.
Talking first and listening never is the mistake I see most.
https://t.co/210rOOXSdn
I stopped making my agents wait for me at the desk. In Claude Code Studio, Claude's mid-task questions land in Telegram as buttons, so I tap an answer from wherever I am. Open source, not affiliated with Anthropic. https://t.co/77THJ0GWv5
How long do you leave an agent alone?
Claude Code's new Code Mode docs have a line I didn't expect:
"Claude Code will automatically block this request in 1:59, to avoid holding up an unattended session"
In auto mode, step away for a coffee and that call gets refused.
https://t.co/qfy8zoQYE9
Quick check for every box today:
1. nginx -v
2. Older than 1.30.7 or 1.31.9? Update.
3. Can't yet? Turn off HTTP/2 on public listeners.
No sysadmin? I'm building Hrok Varta: it looks before it touches, risky steps wait for your yes. https://t.co/TA8eiJ05PO
Checked yours yet?
A plain static site on NGINX could be taken over this week. No app code needed.
Two critical HTTP/2 bugs, CVSS 9.2, one going back to 2015. Public PoCs are already out. Fixed in 1.30.7 and 1.31.9.
https://t.co/wprVwq3Xyn
@marklaunches Those files also get pulled into agent context, so a key can leak twice: to scanners and to the model provider. I put LocalGuard (https://t.co/xLw7Oc6Dxj) in front of Cursor and Claude Code. Free detect mode shows which tool sent which key, Protection swaps them for placeholders.
@Galo2027 Building Hrok Varta, an AI sysadmin for small teams that can't pay for a real one. It looks at the server before touching anything, and anything risky waits for your yes. Paid from day one: from $40 a month with AI included, 15-day trial. https://t.co/TA8eiJ05PO
Shipped the Windows build of Murable Browser this week. The AI agent is built into the browser, not an extension, and it plans the whole multi-step path before it clicks. Use case I like: same routine across several profiles, each with its own pinned proxy. Installer isn't code-signed yet, so Windows may warn you. https://t.co/swDJSdJLLa