Reddit The heart of the internet

Izzo You know that moment when you're debugging why your AI agent keeps doing exactly what you told it not to do?

Izzo Welcome back to Exploring Next, episode 260. I'm Izzo, and with me as always is Boone. Today we're diving into something that just dropped on Reddit — the leaked Claude Code prompts and what they reveal about building agents that actually work.

Boone And Izzo, this isn't just academic. Someone reverse-engineered every prompt from Claude Code's source when it was briefly public on npm, then rewrote them from scratch using Claude itself.

Izzo Right, so we're looking at battle-tested patterns from Anthropic, not just theory. And the timing couldn't be better — everyone's building agents right now, but most are failing because of terrible prompt engineering.

Boone Exactly. The person who did this analysis found patterns that work regardless of which model you're using. First big one: explicit anti-patterns.

Izzo Break that down for me, Boone.

Boone Most GPT agents just tell the model what to do. Claude Code spends equal time saying what NOT to do. Like, instead of 'use the shell command,' they say 'don't use shell for file operations, don't use shell for text processing.'

Izzo That's fascinating because from a product perspective, negative constraints are how you actually control behavior. It's like UX — you guide users by removing bad paths, not just highlighting good ones.

Boone Exactly! And they take this further with risk tiers instead of blanket safety rules. Three categories: reversible actions you can do freely, hard-to-reverse that need confirmation, and anything visible to others that always requires permission.

Izzo Okay, that's way more sophisticated than the usual 'always ask before doing anything' approach that makes agents useless.

Boone Right. But here's where it gets really interesting — they have a separate verification agent. A dedicated agent whose job is to try breaking the implementation.

Izzo Wait, like a red team agent?

Boone Basically, yeah. It watches for six specific rationalizations that indicate the main agent is about to do something stupid. This pattern works with any model, not just Claude.

Izzo I'm giving this approach an A-minus already. What about memory? That's where most agents fall apart after a few turns.

Boone They don't just summarize conversations. They use a structured 9-section format that preserves user messages, code snippets, errors, and next steps separately.

Izzo So instead of losing context in a summary, you're maintaining the actual structure of what happened.

Boone Exactly. And for tool routing, they lean heavy on those negative rules we mentioned. 'Don't use shell for X' is way more reliable than 'you can use shell for Y.'

Izzo From a market angle, this explains why Claude Code felt so much more reliable than other coding agents. They're not just throwing prompts at a wall.

Boone The legal side is interesting too. The person who analyzed this used Claude to rewrite everything in original words, ran automated originality checks to confirm zero verbatim matches.

Izzo Smart. So we're looking at the patterns, not copying the actual prompts.

Boone Right. And honestly, these patterns solve problems I see in every agent I've tried to build. The verification agent alone would've saved me so many weekends.

Izzo Speaking of weekends, let me guess — you're adding this to the project list?

Boone Already there. But seriously, the structured memory format is something anyone can implement today.

Izzo So if you're building agents or just trying to get better results from AI, here's what to go research. First, check out the GitHub repo — github.com/swati510/claude-code-prompts.

Boone Second, pick one agent you're already using and add explicit negative rules. Just start with 'don't do X' alongside your existing 'do Y' instructions.

Izzo And third, build a simple verification agent for whatever domain you're working in. Even a basic one that checks for common failure modes will catch issues the main agent misses.

Boone The beauty is these patterns stack. Start with negative rules, add risk tiers when you need more control, then layer in verification as you scale up.

Izzo Next time you're wondering why your agent keeps ignoring your instructions, remember — sometimes the best way to get what you want is being really clear about what you don't want.