The skills file for claude code have made life very easy. No need to give context manually on the project I am working on. Simply define the project overview in the skills file.
This is basically the anatomy of discipline disguised as tooling. Most people think performance comes from finding a better model, a sharper signal, something new. But the real gains come from constraint—what you’re allowed to do, what you’re not, and how consistently that’s enforced.
Plan mode, permissions, checkpoints… that’s risk management. Context control, compaction… that’s filtering noise. Subagents… that’s knowing when to break a problem into pieces instead of forcing one crowded decision.
It’s the same thing you see in trading desks that actually make money. They don’t rely on brilliance in the moment. They build systems that make it hard to be impulsive and easy to execute what already works.
Left alone, both a model and a trader will drift. Give them structure, and suddenly the same underlying ability looks far more precise.
The CLAUDE.md section hit close to home. Spent about a month treating it like a README - wrote it once, never updated it, and wondered why the agent kept making the same mistakes. Turns out the file drifts fast if you don't update it when something goes wrong.
My current practice: any time I catch the agent doing something wrong, I add a correction to the file immediately. The compaction features you mentioned are also real - once a session runs long enough the context starts leaking and without active compression you get inconsistent behavior. For me the biggest unlock was understanding that the file is the product, not just config.
how do you behave when you have multiple rules and you need only to use 2-3 rules in a session ? if you have a lot of rules claude starts reading those for each new session, while instead I need to specify which rules to use before starting a new session
Awesome explanation on agentic RAG here. +1
The skills file for claude code have made life very easy. No need to give context manually on the project I am working on. Simply define the project overview in the skills file.
This is basically the anatomy of discipline disguised as tooling. Most people think performance comes from finding a better model, a sharper signal, something new. But the real gains come from constraint—what you’re allowed to do, what you’re not, and how consistently that’s enforced.
Plan mode, permissions, checkpoints… that’s risk management. Context control, compaction… that’s filtering noise. Subagents… that’s knowing when to break a problem into pieces instead of forcing one crowded decision.
It’s the same thing you see in trading desks that actually make money. They don’t rely on brilliance in the moment. They build systems that make it hard to be impulsive and easy to execute what already works.
Left alone, both a model and a trader will drift. Give them structure, and suddenly the same underlying ability looks far more precise.
This is great piece of writing for beginners, check out my feed as well!
The CLAUDE.md section hit close to home. Spent about a month treating it like a README - wrote it once, never updated it, and wondered why the agent kept making the same mistakes. Turns out the file drifts fast if you don't update it when something goes wrong.
My current practice: any time I catch the agent doing something wrong, I add a correction to the file immediately. The compaction features you mentioned are also real - once a session runs long enough the context starts leaking and without active compression you get inconsistent behavior. For me the biggest unlock was understanding that the file is the product, not just config.
how do you behave when you have multiple rules and you need only to use 2-3 rules in a session ? if you have a lot of rules claude starts reading those for each new session, while instead I need to specify which rules to use before starting a new session