Muse Spark Code 0.18.0 is out. Several accounts per provider, in the terminal and in ACP editors (Zed, JetBrains, Neovim): save them once, pick one per run. github.com/RandyNorthru...

The best Commander games end with someone at the table saying "wait, read that card again."

A slow query log is the cheapest performance tool there is. Turn it on before you buy a bigger database.

The most useful line in an agent's instructions is usually a ban, not a goal. "Never edit a test to make it pass" does more than three paragraphs about code quality.

Vibe Coded is my Facebook group for people who build open-source software with AI coding tools. Beginners are welcome, nobody gets mocked for using AI, and anything you share needs a public repo and a license. Come say hi: www.facebook.com/groups/vibed

A counterspell works because Magic's stack is last in, first out: your response resolves before the spell it answers. Learning to hold priority is basically learning to read a call stack.

"Works on my machine" got harder with AI, not easier. The agent's machine is now a third machine nobody looked at.

A backup you've never restored from is a hope, not a backup. Test the restore.

A test that has never failed hasn't proven anything. Plant a bug, watch it go red, restore the code. Then trust it.

Lumafoil watermarks photos, PDFs and video in the browser: text, logos, signatures and QR codes as reusable presets, batch processing, and image metadata stripped by default.

Building a 60-card Magic deck is dependency management. Every slot has to justify itself, and one bad include breaks the whole build.

Second opinions for agents: a Claude Code skill that asks Codex to review the code read-only, then Claude checks every finding against the code before reporting it.

"The Wi-Fi is slow" is often a crowded 2.4 GHz channel. Check what channel you're on before buying a new router.

Handing a project from one AI agent to another, or one machine to another, loses context. My teamwork skill writes a hash-checked handoff the next agent resumes from, and it refuses to carry likely secrets.

If your agent can push to main, your agent is your release process. Is that the release process you want?

Tip for anyone automating Windows: UI Automation first, OCR as the fallback, screen coordinates last. Coordinates break on the first DPI change.

Prompt injection is SQL injection with better grammar. Every web page your agent reads is untrusted input.

Baffleword, my word-game parody: 191,745 tournament words in a trie, a solver that finds every word on the board, and 6x6 boards with blocked tiles and "TH" dice. Built with .NET 10 and Avalonia.

Learning Nmap without pointing it at someone else's network: 358 scan profiles plus a local lab on 127.0.0.1 with fake services to scan. Public targets need an explicit flag, and every profile has a dry run.

Open the Windows: 458 Windows 11 tweaks, but only ones verified on 25H2 apply by default. Every change is journaled and reversible, and it will never turn off Defender or UAC.

click_text in my Windows MCP server finds text on screen and clicks it: UI Automation first, OCR fallback, no coordinates. 53 tools for Windows agents, with read-only profiles and dry runs.

MCP is quietly becoming the USB port for AI tools. One protocol, and an agent can drive a browser, a desktop or a database. The hard part now is permissions, not plumbing.

Pokémon starters as coding styles: Bulbasaur writes tests first. Charmander ships fast and sets things on fire. Squirtle refactors everything until it's calm. Which one are you?

Question for people running local models: what's the smallest model you'd actually trust for code review, not just autocomplete?

Every tool an agent can call should have a dry-run mode. If you can't preview what it's about to do, you can't honestly approve it.

A raymarched 3D Mandelbulb in about 5.5 KiB of gzipped JavaScript. WebGL2 does the heavy lifting; the budget was 80 KiB. Five palettes, and power and iterations change live. randynorthrup.github.io/mandelbulb-e...

Agents should fail loud. A quiet "done" with a skipped test is worse than a red build, because you'll trust it.

Magic: The Gathering is Turing complete. Researchers built a working Turing machine out of real cards, which means deciding who wins a game can be undecidable.

Before blaming the code: DNS answer, stale cache, cert chain, MTU over the VPN, and whether you're on IPv6 while the server isn't. What's on your checklist?

S.A.K. Utility: a portable Windows toolkit for PC techs. A partition manager that queues changes before applying, PST/OST tools with no Outlook, a vulnerability scanner fed by CISA KEV, NVD and OSV, and an AI tab with 25 technician workflows. sakutility.com