Jorge Mena (@JorgeMenaDev) header
Jorge Mena (@JorgeMenaDev) profile picture

Chile πŸ‡¨πŸ‡± β†’ London πŸ‡¬πŸ‡§. Dad of 2. Senior SWE building in public. 🐦 https://teapaw.org/t.co/EmQMsroqF4 $68/mo πŸ›‘οΈ https://teapaw.org/t.co/FU1a4HkjIH πŸ—» https://teapaw.org/t.co/3RG7m6uVIg

London
Joined December 2010
Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)Photo by Jorge Mena (@JorgeMenaDev)
It took me 7 years to finish a 5 year business degree in Chile. Then I couldn't find a job, so I sold toys in my best friend's shop in Valdivia while we saved to go to Australia. I got there at 27 with almost no English. Made pallets, picked avocados, met my wife. I learned to code at 30 during lockdown. Now I'm an AI engineer in London with a day job at a bank, two companies and two kids under 2. It isn't too late.
66
77
7
1,532
64,024
Ninja ice cream machine please never leave me I love you 🫢
62
My agent's report said one of my products got 125 sessions from Chile in 28 days. The real number was 580. Nothing was broken. Since mid October, cookie-free visits store the country in a different field. The report read the old one, so 604 sessions had no country at all. The agent that wrote the report missed it. A reviewer agent caught it. If an agent hands you a number that looks low, ask which field it read.
72
Andy's site showed up 26,483 times in Google in October. More than double September. Clicks: 99. September was 101. The biggest share came from one Spanish blog post about using WhatsApp Business and normal WhatsApp on the same phone. 9,306 impressions, 24 clicks. So more pages got me seen, not visited. The part I'm losing people on is what they read in the results.
45
31 chat button designs rejected in one night. I wanted a new launcher for Andy's chat widget, so I put two AI models on it as designers. 11 runs, 6 rounds, 33 designs. I kept 2 as maybes. Then I looked again and typed "don't like any. keep trying". Making designs costs nothing now. Knowing which one is right is still my job, and I haven't found it yet.
52
My scheduled agent jobs all stopped after a T3 Code update. The cause: login now downloaded a 110 MB file first, and the scheduler gave each job 20 seconds. The fix: Finish the download before login. Show the real error, not a quiet stop. If your scheduled jobs go silent after an update, check the timeouts first.
2
1
58
Here's mine. I told claude to use codex cli to create the image πŸ˜‚
In 2 months it’ll be 2027. And we still don’t have image generation in Claude.
1
2
112
4 review rounds in a row found the same kind of bug. I built a small tool to handle replies to my Andy cold emails. When someone replies, it shows a button so I know I still owe them an answer. Each review found a new way that button could disappear before anyone had answered. My agents patched it. The next review found another one. The patches weren't the problem. The tool decided what was answered by comparing timestamps. So we stopped patching. Each message now records its own answer, and the timestamp logic is deleted. No real reply ever hit the bug. If your fixes keep coming back as new bugs, stop patching and change the idea underneath.
2
52
I was paying $99 a year for a workout app. So I built my own. It's called Fitbite, and it's just for me. It started with recipes. I keep the ones I cook most, see their macros, and tap one button to get a shopping list. If two recipes need onions, it adds them up. I share the list with my wife or take it to the supermarket myself. Then I added my workouts and cancelled the subscription. Newest feature: the machine you need is taken halfway through a workout, so you swap the exercise right there. The new one comes up with the weights you lifted on it last time. When you can build the app, you stop settling for someone else's.
3
1
585
I've been testing cold outreach for Andy for a while now. At first most of my emails went straight to spam. So I kept trying different approaches. 183 emails later: 2 replies. 1 positive. The first 93 got nothing. Now I'm a bit stuck. Outreach is the part I can't crack yet. If you've made cold email work for a small product, what changed it for you?
6
2
141
I have literally no time to play FPL this season. Two kids, a 9 to 5, and Andy at night. But I'm not losing my work league to my friends again πŸ˜‚ So I built an agent skill that does the research for me. Before each deadline it reads the transcripts from five YouTube channels I picked, checks injuries, fixtures and odds, and gives me three transfer options with the trade-offs. Same setup I use for work, pointed at my weekend hobby.
1
65
32 people opened my app. 11 saw the paywall. 0 started a trial. My Wainwrights agent pulled that funnel on its own, with a read-only PostHog key for one project. I never opened a dashboard. The leak is obvious now.
2
1
135
Day 1 of building Andy in public. My raw post ideas were all over the place, so I built an Ideas board into Andy, the scheduler I use every day: - paste the note as it came out, even a voice note - my agents save ideas there too - turn one into a draft when it's ready Day 2 tomorrow.
Do you put your mrr on @trust_mrr at $68, or wait until it looks good? I listed Andy today. Hiding it wasn't making it bigger. Day 0, game on. let's do it!!!!!!!!! πŸ’ͺ
1
4
148
Do you put your mrr on @trust_mrr at $68, or wait until it looks good? I listed Andy today. Hiding it wasn't making it bigger. Day 0, game on. let's do it!!!!!!!!! πŸ’ͺ
1
3
960
Anthropic's new guide says fold most subagents back into one agent that loads skills when it needs them. Keep subagents for checking work, parallel jobs and context that won't fit. I run a crew of agents. I think they're mostly right. Where the crew cost me: Three parallel agents stopped mid-build when I hit a session limit. Uncapped parallel phone builds pushed my 10-core Mac mini to a load of 173 to 209. A handoff bug stranded 1 of 5 handoffs between agents. One agent never pays that. Where it still pays: A second model checking the first. In about 12 of 16 sampled cases, the reviewer corrected the main agent's claim or design. Real parallel work. Seven agents painted one chapter each of a 93.5 s music video. Both are on Anthropic's own keep list. My rule now: split by context, not by job title. If an agent only exists to pass a summary along, it should be a skill. How many of your subagents would survive that?
4
6
1,039
Running a crew of agents? They need somewhere to leave the decisions that are waiting on you. I use Paperclip for that. A self-hosted orchestrator that treats your agents like a small company. βœ“ MIT, runs on your own machine βœ“ a decisions inbox and tickets, so nothing sits in a chat scroll βœ“ triage passes on a schedule Paperclip github.com/paperclipai/paper…
1
4
657
Weird one with Andy, my AI that answers customer messages. A bug disconnected a customer's WhatsApp from Andy. A message came in anyway and got an AI reply, tagged "IA". It was Meta's own AI, answering on their WhatsApp while Andy was down. The channel is now a competitor.
1
2
477
Keyword research for 9 cents. Before building a nearby fell finder for my Wainwrights app, I checked 59 phrases with DataForSEO. "Near" searches came to about 530 a month. So I shipped it. βœ“ $0.09 for the whole run βœ“ "wainwrights" alone: 14,800 UK searches a month βœ“ my Cursor project runs it next to Firecrawl and Search Console DataForSEO dataforseo.com
3
1
365
Your agents keep working on one machine while you check in from another. T3 Code is where my agent crew runs. It lives on my Mac mini. I start threads from my phone. βœ“ open source βœ“ runs Codex and Claude agents side by side βœ“ one thread hands work to child agents and hears back when they finish T3 Code github.com/pingdotgg/t3code
3
1
250
Software isn't dead. Writing it stopped being the hard part. My agents write the features now. The part of my job that got bigger is throwing things away. Two models built the same landing page from one brief. I kept one. Three rounds of icon options for Andy. Rejected all three. Cheap to build. Still expensive to choose.
2
1
161