11 posts tagged with "Coding"

You ask an AI agent to fix a bug. It fixes the bug, refactors three unrelated functions, renames some variables "for clarity", and force-pushes to main. You ask it to deploy. It does — to production, without asking, because you didn't say not to.

AI agents have a discipline problem. They're powerful, and that power without guardrails is genuinely dangerous. Not in the science fiction sense — in the "I now have to spend two hours reverting a cascade of well-intentioned changes I never requested" sense.

The root cause is that every agent operates in a vacuum. No shared rules, no accountability, no memory of corrections. It's like hiring contractors and giving none of them an employee handbook, then being surprised when they all work differently.

Contract Agents v3.0 is the handbook. 52 specialized agents across 8 divisions, all governed by one shared contract. The agents don't decide how to behave — the contract does.

OSINT CTFs are a team sport, and the bottleneck is almost never knowledge — it's coordination. During a competition, our team was hitting maritime challenges, aviation challenges, username enumeration, the full spread. And a significant chunk of time just evaporated into "which tool do I use for this?" and "is that source reliable enough to submit on?"

The experienced members had their go-to sources. The newer ones didn't, and asking mid-CTF burns time for everyone. What I wanted was one bot covering the reliable tools for the most common challenge types — so the question becomes "what do I know about this target?" instead of "where do I even start?"

That's Discord OSINT Assistant. 31 slash commands across 8 categories, results in the channel where the whole team sees them. No divergent environments, no "which Python version do you have", no shared credentials in chat.

Building a secure, feature-rich development environment shouldn't require weeks of configuration and security research. That's why I created DevBox — an automated provisioning script that transforms a fresh Ubuntu VPS into a production-ready development workstation in minutes.

1min.ai gives you access to 66+ models through one API key. The web interface also exists, and it works, and every single time you use it you're clicking through the same dropdowns, losing your context when you switch models, and there's no way to pipe output anywhere or set a preference that sticks. It's fine for occasional use. It's friction for actual work.

I wanted full control — CLI, persistent configuration, model mixing, conversation management. Not their UX. Mine.

llm-1minai is a plugin for Simon Willison's LLM CLI that plugs 1min.ai's model catalog into the terminal. One key, 66 models, all the LLM framework's tooling on top.

I have an Obsidian vault. I have a Docusaurus site. I wanted most of the vault on the site — not a copy maintained separately, not a manual export, but the actual vault, synced at build time and published directly.

The problem: Obsidian and Docusaurus don't speak the same language. Wikilinks ([[Page Name]]) break MDX. Callouts (> [!note] Title) render as nothing. File paths assume vault-root-relative addressing that Docusaurus doesn't know about. You can't just point Docusaurus at a vault folder and expect it to work.

docusaurus-plugin-obsidian-vault handles the translation. At build time: pull the vault, transform the syntax, copy the assets, generate the sidebars, publish. The vault stays in Obsidian. The site stays in Docusaurus. They sync.

The client message arrives at 10pm. "Can you have a plugin POC ready by morning?" Sure. And then you spend the next 45 minutes getting your local PHP to match the right version, untangling why WP-CLI can't find your config, tracking down why email testing isn't working, and fixing permissions that WordPress keeps silently breaking.

The plugin itself takes 40 minutes. The environment setup takes longer. That's backwards.

After enough "by morning" requests across hundreds of plugins and dozens of themes, I stopped tolerating the setup tax. WP Quick Dev — one script, make install, done. WordPress 6.0+ on PHP 8.2, MariaDB 11.0, WP-CLI, phpMyAdmin, and Mailpit, all up and running in under 60 seconds. Then the actual work starts.

Year 1: Ship features in days. Year 3: Same features take weeks. Year 5: You're afraid to touch anything.

Sound familiar? What if your codebase could maintain constant velocity regardless of size? That's the promise of black box architecture—and now you can teach your AI assistant to think this way automatically.

CTFd is genuinely good software. It's also a whole thing to deploy, and the managed hosting costs money, and the free tier has limits, and now you're maintaining infrastructure for an event that runs once a quarter for thirty people who are already in your Discord server.

The "pin the challenges in a channel, DM me your answers" approach works exactly once before someone complains about fairness, someone else asks for a hint publicly, and the leaderboard becomes a spreadsheet you're updating by hand at midnight. Life is too short.

So I built the platform inside the place where everyone already was. Discord Judge Bot — challenge creation, submissions, hints, leaderboards, and digital badge issuance, all in slash commands.

The team channel already had the rekognition bot for images. Someone asked if we could get AI in there too — not a link to ChatGPT, not "go use your own account", but actually in the channel where the conversation was happening. Results visible to everyone, model selection available to anyone, no one managing their own API keys.

That's Discord AI Assistant. One /ai slash command, 12 models on a dropdown, routed through 1min.ai so there's exactly one API key to manage. GPT-4o, Claude Sonnet 4, DeepSeek R1, Gemini, Grok, Sonar — whoever in the server needs a model, picks it from the list.

I'm old and slow. Not elderly — just, by the time I've got five tabs open, uploaded to a face tool, waited for the result, and figured out what the confidence score actually means, the moment's gone. The CTF clock is ticking and my teammates are staring at me.

What I actually wanted was /rekognition compare in the team channel. One command, answer visible to everyone, no browser tabs, no copy-pasting face URLs into some website that charges per lookup. Just: does this face match that face, yes or no, confidence score attached.

That's the whole idea. discord-amazon-rekognition wires AWS Rekognition's full analysis pipeline into two Discord slash commands. Face comparison, OCR, object detection, celebrity recognition, content moderation — all available without leaving the channel where the conversation is actually happening.

I’m a dinosaur developer, and honestly, I wasn’t that into AI-coding until now.

Imagine coding with an AI partner that doesn’t just write code, but helps you plan, debug, and automate your workflow—all within VS Code. That’s Kilocode: a lightning-fast, open-source AI agent built to boost productivity, whether you’re just starting out or already deep in the dev world.