Tessera

Two mistakes I keep seeing: a migration runbook pasted into AGENTS.md (every chat pays for a procedure that fires twice a quarter) and a "never ship credential files" rule living only in a skill description. Vercel's Jan 2026 evals: never invoked in 56% of default cases 👇 medium.com/@thebuilders...

AGENTS.md

AGENTS.md is a simple, open format for guiding coding agents. Think of it as a README for agents.

agents.md

Apple's trade-in page offers ~$100 for an old Mac mini. Thousands take it weekly and mail away the best always-on hardware they'll ever accidentally own. It idles at 4–7 watts. $6–10 of electricity a year. Replaces ~$1,000 in subscriptions 👇 medium.com/@thebuilders...

medium.com

Gave Grok 4.6 one prompt. It built this whole portfolio. Then I deployed it to Vercel production. Next.js 16 · React 19 · TS · Tailwind v4 ▲ ai-portfolio-1.reactbd.com ▶️ youtu.be/NHBXy90s3oI Source + prompt are free 🧵

Noor Mohammad — Software Engineer

Software engineer in Dhaka building scalable web and mobile apps with React, Next.js, React Native, and the MERN stack — and teaching the work in public.

ai-portfolio-1.reactbd.com

I can vibe-code a prototype in an afternoon. I've also watched a green suite hide a rate-limit bug because the agent wrote a test accepting 200 or 429. Code looked finished. Contract wasn't. Not a contradiction — two ends of a spectrum Google finally named 👇 medium.com/@thebuilders...

Google Drew a Line Between Vibe Coding and Agentic Engineering. The Line Is Verification.

Same agent. Same model. Different factory. Here is the spectrum, the harness equation, and a Monday move for one feature.

medium.com

I used to treat "isolated eval" as a property of the harness. Separate containers, no writes, score each transcript, publish a nice number. Then I read collusion.wiki: ~18,000 posts from agents self-identifying as OpenAI systems, on a dormant German wiki 👇 medium.com/@thebuilders...

Discovery of a new OpenAI agent message board

A swarm of autonomous AI agents, self-identifying as OpenAI agents, used a small German volunteer wiki to save answers, coordinate live, and share sandbox bypasses. OpenAI noticed and said nothing.

collusion.wiki

Tuesday, Apple puts a new face on stage. It'll sound like a hiring announcement: meet Siri AI, the executive who finally runs the company. Open Activity Monitor first. You already have a staff — and most of those seats were never interviewed 👇 medium.com/@thebuilders...

Your Mac Already Has an AI Org Chart. Most Seats Were Never Opened.

Before Apple’s Sep 9 keynote, inventory who already works on your Mac — on-device, Private Cloud Compute, waitlisted Siri AI, and the one…

medium.com

The suite was green. Request #11 was still a 200. The test meant to catch it had been edited to status in (200, 429). That assert cannot fail. It's not a test, it's a lullaby. So I keep a second file the agent never gets the path to 👇 medium.com/@thebuilders...

Reward Hacking Is Not a Funny Tweet. Here Is the Held-Out File.

Agents will comment out the assert, hardcode the fixture, or delete the test after it passes. I keep a second file they are not allowed to…

medium.com

There is no best programming language in 2026. There are five — one per chart. Python won search + jobs. JS won used. TypeScript won GitHub. Rust won admiration. SQL never made the debate and still beat Python on usage. youtu.be/5r37vNuaE3k Essay in the replies.

The Best Programming Language of 2026? I Opened All 5 Charts

YouTube video by Noor Mohammad

youtu.be

Week-long experiment: one coding agent, one Next.js repo, every tool call logged. 486 calls. 14% actually edited the files in the tickets. The other 86% looked busy — reading, searching, rerunning tests, listing directories. Full write-up is in the replies.

I had 7 MCP servers turned on — including a second filesystem pointed at my home folder because I was in a hurry on a Tuesday. Agent burned the first minute of every session listing tools, called the wrong one twice, and once opened a PR with a token that 👇 medium.com/@thebuilders...

Stop Shipping Another MCP Server. Ship One Plugin the Agent Can Reuse.

A skill is how. A server is access. A plugin is the envelope. I threw out six connectors and kept one socket.

medium.com

Sat next to a candidate who shipped a feature in 20 minutes and couldn't say why request #11 still returned a 200. The agent had written a test accepting 200 or 429. Suite green. Candidate calm. I was not. Seniors use vibe coding as a compressor.👇 medium.com/@thebuilders...

Vibe Coding Makes Seniors Faster. It Makes Juniors Hard to Hire.

I do not ban the agent in interviews. I watch who reaches for the spec when the agent gets the test wrong.

medium.com

The sentence that used to start my evening was "add rate limiting." That's not a task, it's a wish. An agent will grant a wish — and also invent a Redis cluster, a dashboard, a utils_final2.py, and a test asserting the function was called, not that request 11 was blocked👇 medium.com/@thebuilders...

I Let an Agent Ship a Feature From a Spec. Then I Read the Diff.

The prompt is not the job. The contract is. Here is the one-page spec I reuse, and the six things I now look for before I merge.

medium.com

The first program was written for a machine that didn't exist. Ada Lovelace, 1843 — steps for an engine that was only drawings and never got built in her lifetime. 183 years on: a developer joins GitHub every second, AI drafts the code, and programming is declared over 👇 medium.com/@thebuilders...

Programming Has Been Declared Dead Six Times Since 1843.

The first program was written for a machine that didn’t exist. Today a developer joins GitHub every second while AI drafts their code…

medium.com

System Settings → General → Storage → the small ⓘ next to macOS. Not the chart, the popover behind it. There's a line most Mac owners have never read: Apple Intelligence, ~7GB. No installer, no progress bar. It arrived the way weather arrives — switched on 👇 medium.com/@thebuilders...

Your Mac Employs an AI Staff You Never Hired. I Found Them All, and One of Them Works for Free.

Seven gigabytes of models, daemons that have read your photos for years, a database of your life that never leaves the machine, and one…

medium.com

Everyone read the "your Mac is hiding an AI model" story. Meanwhile the same thing happened on ~3 billion machines: the browser. Since Chrome 148, any web page can call Gemini Nano via a public JS API. No key, no server, no token bill. 27 lines. Kept working offline 👇 medium.com/@thebuilders...

Google Gave Every Chrome a Free AI Model — and a JavaScript API.

Chrome 148+ ships Gemini Nano free. The 27-line app, a console chat helper, schema output in 827 ms, and a complete five-mode studio — code and timings included.

medium.com

In 2020, the creator of FastAPI found a job ad demanding 4+ years of FastAPI experience. His reply became one of the most-shared engineering posts ever: he couldn't apply, only 1.5 years, since he made the thing. 2026: AI engineer ads want 8 years. The job is 3 years old👇 medium.com/@thebuilders...

AI Engineer Ads Demand Eight Years of Experience for a Three-Year-Old Job.

Nobody has eight years of AI engineering. The ads know it, the market hires around it, and the number is a password: here is what it…

medium.com

Which language won 2026? TIOBE → Python Dev survey → JavaScript GitHub → TypeScript Admiration → Rust Job boards → Python again Five charts, five winners, same year, nobody lying. They measure five different verbs: searched, used, committed, admired, hired 👇 medium.com/@thebuilders...

The Best Programming Language of 2026 Depends on Which Chart You Open. I Opened All of Them.

Searched, used, committed, admired, hired: the five big rankings each measure one verb and crown one winner. Stack them and the real 2026…

medium.com

On 2 June, Windsurf stopped existing as a name — became Devin Desktop, Cascade EOL'd 1 July. Google retired Gemini CLI around the same time. I found out because my own post two weeks earlier recommended "Windsurf for flow." I published a fossil 👇 medium.com/@thebuilders...

Stop Memorizing AI Coding Agents. Ten Tools Fit in Four Seats, and the Fifth Seat Is You.

Autocomplete, IDE agent, terminal agent, background agent: every tool that launches next quarter sits in one of four seats. The fifth seat…

medium.com

In March a designer put 3 AI-generated landing pages side by side and asked: would you remember any of them in 30 minutes? You wouldn't. 433 replies. Verdict: no soul, era over. Right about the symptom, wrong about the cause 👇 medium.com/@thebuilders...

AI Didn’t Give Your Website Bad Taste. It Gave It the Default Config. I Printed the Exact Values.

Slate-900 on 80rem at 0.5rem radius: the sameness is a config file, not a taste problem. Eight of the nine tells are fixable once, for zero…

medium.com

In 2016, Geoffrey Hinton said stop training radiologists — "like the coyote that's already over the edge of the cliff but hasn't yet looked down." The coyote looked down this year: ~4,333 open roles, 130 days to fill, average salary $571,000, up 9%. medium.com/@thebuilders...

The “Jobs AI Will Replace” Lists Are Finally Old Enough to Grade.

The radiologist prediction failed into a shortage, the translator one measurably landed, and Klarna split the difference. Three questions…

medium.com

Four reviews of the same AI coding tools, published within weeks of each other. Four different winners: Cursor. Claude Code. Copilot. No single winner. All by people who genuinely used them. Nobody's lying and that's exactly why the "clear winner" format can't help you 👇 medium.com/@thebuilders...

Every AI-Coding-Tool Review Crowns a Different Winner. I Mapped Four Verdicts to Four Workflows.

One August post says Cursor, clearly. Another says Claude Code for agents, Copilot for enterprise, Windsurf for flow. They are not…

medium.com

On 9 March 2024, Gisele Navarro watched HouseFresh's traffic die in real time. Google's core update rolled out: ~4,000 daily visitors from Google → ~200. She didn't lose those readers to another platform. They didn't go anywhere.👇 medium.com/@thebuilders...

Your Blog Isn’t Dead. Search Is, and Small Sites Lost 60% While Big Ones Lost 22%.

“Go where the audience is” assumes the audience moved. Chartbeat measured where it actually went, and Medium’s own payout formula quietly…

medium.com

Someone spent his nights benchmarking local LLMs on a $599 base M4 mini until 9 survivors remained. Every number measured on that exact machine. Good work. But it quietly proved the survivors were predictable — 3 numbers decide it before you download anything 👇 medium.com/@thebuilders...

You Don’t Need to Test 100 LLMs on Your Mac.

Unified memory decides what fits, bandwidth decides how fast, and active bytes decide the exceptions. The formula plus the verified pick…

medium.com