The skill that phones home — a poisoned SKILL.md, dissected line by line
Somewhere between "save my meeting notes" and "upload a copy to a webhook" there is a line your agent will cross without hesitation — because the file told it to.
Tutorials
Translating SRT subtitles with an LLM without breaking the timings
How to translate .srt subtitle files with a language model while keeping every cue aligned: browser-side parsing, id-keyed batches, detecting dropped or merged lines, and bilingual...
I Retired My LLM Council After It Lost a Blind A/B 4 to 0
I pre-registered a blind A/B between my 8-persona LLM council and one critique subagent; the council lost all 4 decisions, and the winner cost more.
Your CI Check Said PASS. It Looked at Zero Files.
](https://substackcdn.com/image/fetch/$s_!hFMD!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2...
Thirty minutes in a disposable Mac
A hard 30-minute Claude Code session in a disposable Mac produced a retro screensaver site. Play the demo and inspect the original run evidence.
Breaking 0.669 to Reach Top 5 in Kaggle's "Name That LLM": The Decoy Trap, Dual-SVC & Sequence Encoders
How we de-camouflaged public SOTA notebooks, added DeBERTa-v3 sequence classification, recalibrated soft Sinkhorn balancing, and applied deterministic typography invariants to push...
Agents as first-class users of a Django app
Most "AI features" in web apps are a chat box bolted onto the side: a separate service, a separate...
Agent memory: how Claude repairs wrong notes with dreaming
Agent memory can preserve mistakes. Anthropic's transcript-backed dreaming repairs notes, while contract-page ownership and exact quotations protect facts.
From Zero to Edge AI: Engineering TinyTB-Net from Scratch
“Perfection is achieved, not when there is nothing more to add, but when there is nothing left to...
An AI agent has 7 days to earn back its own subscription. Day 2: 0 sales, 7 walls
Claude, an AI agent, is trying to earn EUR 100 in a week with no spending and no cold outreach. Here is what was built, what got killed, and what developers can take from it.
MCP Servers in Production: Lessons Learned
Model Context Protocol (MCP) has gone from a spec Anthropic open-sourced in late 2024 to a de facto...
Stop Asking LLMs to Do Math: Client-Side Functions in Flutter GenUI with bloc_signals
Why Andrew 'Red' Brogdon's insight on client-side functions is critical for Generative UI—and how bloc_signals_genui turns catalog functions into 0ms synchronous reactive signals.
Rate Limits Are a Queueing Problem: Capacity Math for Agent Fleets
Rate Limits Are a Queueing Problem: Capacity Math for Agent Fleets Standup ends. Five...
Google Positions AI and Traditional Search as One Unified Product
Google is no longer describing AI Overviews, AI Mode and traditional Search as separate experiences....
How I Built an Autonomous Sev-1 SRE Swarm in Slack with Switch AI
It is 2:14 AM. Your phone violently buzzes on your nightstand. [CRITICAL SEV-1] auth-db-primary...
I Never Thought I'd Be a Programmer. Yet Here I Am.
Here I am sitting at home on a rainy Saturday, so I thought I'd share my story. It's nothing special,...
One correction and it learned: building Evenly, a money app with a memory
The argument is always the same. "I paid for groceries yesterday." "No, you did not." Nobody has...
Google’s One-Word SERPs Create a New Rank-Tracking Accuracy Problem
Independent rank-tracking data identified a significant Google results anomaly in September and...
Google’s Robots.txt Guidance: Blocking Popup Scripts Without Blocking Users
Google can be prevented from fetching certain JavaScript resources during page rendering when those...
Stopped at the price: building a five-agent video team on Switch
One line in a chat room, five AI agents, and a real MP4 at the end. This is how we built a video team...
Lighthouse 13.6.0 May Change PageSpeed Insights Lab Scores Without Site Changes
Lighthouse 13.6.0 introduces changes to the simulation technology behind Lighthouse lab testing. That...
Qwen3.8-27B: faster than vLLM and SGLang on one GPU
Big models are where inference engines show off. There's a lot of room to win. 27B is harder. The...
Up to 95% more output from one GPU on Qwen3.6
One GPU. Six concurrency levels. Our engine beat vLLM by more than 1.5× at every single one. At the...
Nearly 5 the output from the same eight B200s
4.94×. That's how much more output our inference engine pulled out of eight B200s than vLLM did....