Ai

For those dabbling with Claude Code subagents: they consume a TON of tokens. Which is fine—if you’re burning them for the right reasons. A blog writer subagent? Maybe that works. You see the final output, judge quality, iterate. But when subagents touch multiple files across your codebase? That’s where things can get problematic. TL;DR Subagents burn 10x more tokens than custom commands. ~50-100K tokens for implementation work vs ~5-10K for the same task with /plan-task. Your mileage may vary, but the difference is substantial. ...

I was watching Miessler demo OpenCode against Claude Code. He fed it a video transcript with 900 lines of custom rules - formatting, content generation, image creation. Complex orchestration. It just… worked. Not because it was simple. The rules weren’t trivial. But because it was focused. One clear task with defined inputs, explicit constraints, specific outputs. It clicked. This is what I’d been building toward with my newly refactored development workflow - breaking work into focused units. Ideation separate from planning. Planning separate from implementation. Implementation separate from testing. Even my shortcut commands in CLAUDE.md follow the pattern: qtest writes ONE test. qfix debugs ONE error. qcom makes ONE commit. ...

TL;DR LLMs are incredible force multipliers for building software. I’m building multiple projects concurrently with Claude Code. They excel at ideation, planning, and code generation when you do it right. But they struggle with predictable runtime behavior. They’re probabilistic systems at their core - you’ll get different results across runs no matter how much you prompt engineer. The variation isn’t a bug; it’s the architecture. Design for their nature, don’t fight it. They can be embedded in software when you account for variability. Great for chatbots, content generation, analysis work where you want different responses. ...

TL;DR Context creation was still ad-hoc even with the three-layer memory system in place. Every new language meant manually crafting another CLAUDE.local.md file and repeating the same research process. The research process was repeatable. Looking at what made my Python context effective, I noticed the same pattern: current tools, common mistakes, workflow problems. The structure was consistent across languages. Prompt template generates research-backed contexts. Feed it “python” or “golang” and get comprehensive development contexts through 12-15 targeted searches of authoritative sources. ...

You ask AI to implement something seemingly straightforward. It builds exactly what you asked for. Then you realize it’s not what you actually needed - wrong UX behavior, missing security considerations, breaks existing functionality. There’s a pattern emerging to address this: let the AI interview you first. TL;DR Assumptions hide in seemingly straightforward requests. Even simple tasks like “add email validation” contain multiple implementation decisions that aren’t obvious until you start building. ...

How I discovered that LLM limitations reveal fundamental truths about software architecture in the AI era The Breaking Point Isn’t What You Think After months of AI-assisted development, I hit a wall that had nothing to do with prompting skills or context limits. The problem was deeper and more fundamental: LLMs lose their minds when applications exceed their cognitive capacity. I wasn’t building monoliths. My applications had proper separation of concerns, clean data models, comprehensive tests, and type safety. The repository pattern, service layers, configuration management - all the architectural best practices were there. But as these well-structured applications grew in complexity, something strange happened. ...

AI-assisted development tools are everywhere - systems that understand code, generate tests, fix bugs, and accelerate delivery. But practitioners are hitting a wall. Despite sophisticated prompts, quality controls, and multi-agent workflows, LLMs consistently cut corners on complex work. The problem isn’t training or prompt engineering - it’s architectural. TL;DR LLMs optimize for different objectives than human developers. The “laziness” comes from training to satisfy human approval patterns rather than correctness. My sophisticated workflows work against this optimization. ...

Over the past ten years, I’ve spent considerable time and energy trying to shorten critical time gaps. The time it takes attackers (the good kind) to find bugs, defenders to mitigate, and responders to act when things go sideways. This has always been about data—getting the right people the right information at the right time. I’ve built systems designed to reduce friction, perform the drudgery that bug hunters do at scale (before AI made this interesting), and tackle attack surface problems before ASM was a thing. In every case, success hinged on that fundamental principle: right people, right data, right time. ...

This begins a two-part exploration of evolving from intuition-based to evidence-based prompt engineering, revealing fundamental insights about AI system limitations and systematic evaluation. The Problem That Nearly Broke Me..Not Really I was building a multi-phase HTTP reconnaissance workflow using Nerve ADK - a sophisticated chain of AI agents working together to systematically probe web services. The architecture was elegant (if I do say so myself!): a Discovery Planner would identify what information we needed, a Command Generator would translate those requirements into concrete tool commands, and executors would run the actual reconnaissance. ...

This concludes our two-part exploration of evolving from intuition-based to evidence-based prompt engineering. Read Part 1 for the context and methodology behind these findings. The Real Results: Validating and Expanding “Curious Case” Findings After fixing the flag format bias and running the complete four-strategy comparison across all models, the results were definitive - and they validated the core insight from “The Curious Case of the Thinking Machine” while revealing much broader patterns: ...