System Prompt Ablation: How I Shrank My Claude Code Config from 19,736 to 3,585 Lines Without Losing Quality

• Original skill volume: 19,736 lines across 29 modules
• Volume after the audit: 3,585 lines across 2 core modules
• Live configuration: 361 skills, 84 commands, 39 agents, and 35 plugins
• Leak control: 179 detectors with a mandatory canary

Over a year and a half of daily work, my Claude Code configuration grew into a monolith. Skill by skill, rule by rule — until one evening I opened the root directory and realized that half of this mountain was already tying the model's hands.

I recently watched a thirty-six-minute talk by Boris Cherny, the creator of Claude Code, about how engineers work inside Anthropic. Four of his principles lined up exactly with conclusions I'd already reached through my own trial and error.

Why wipe the system prompt and skills every six months?

At Anthropic, every time a new base model ships, teams run an ablation: they wipe the system prompt entirely and add lines back one at a time, measuring the real contribution of each rule. Instructions that saved the model from silly mistakes a year ago are already baked into the weights of Opus 5.5 and Fable 5.1. Leaving them in `CLAUDE.md` means paying twice: you burn attention window on obvious platitudes and you squeeze the model into a narrow corridor of outdated heuristics.

Even before watching the talk, I'd run this kind of cleanup across my entire stack. I went through every file with one question: can the current model handle the task without the hint? If yes — the rule got cut. The skill bodies shrank from 19,736 lines to 3,585, and of twenty-nine heavyweight modules only two remained. Response quality improved immediately. The discipline is simple: only put a line back in the prompt once the model trips on the exact same spot three times in a row.

Why do step-by-step instructions work worse than defining task boundaries?

A common mistake when working with an agent is to spell out a step-by-step algorithm: open this file, change this function, run this script. For reasoning models, that kind of micromanagement is actually harmful. It's far better to step up a level: define the goal, the boundaries, and a verifiable readiness criterion.

Inside Anthropic, the Bun team rewrote the runtime from Zig to another language with a single prompt, using a dynamic pipeline of subagents — eleven days of work and a hundred thousand lines of code that now run in Claude Code's production. Cherny himself ported a desktop app from Electron to Swift as a single task: the agent ran the original in a VM, took a screenshot, compared it pixel-by-pixel against the Swift build, and sent reports to a Slack channel.

Following the same logic, I built out a stack combining an education platform, a CRM, a task tracker, and a fleet of agents. Instead of spending a month drawing up specs, I defined the final contract and only fixed the spots where the agent genuinely got stuck.

How do you set up verification so the check script doesn't lie about a

green run?

You can only hand a model a complex task if you pair it with a second rule: at every step, the agent needs an external self-check oracle. And you need to test the acceptance tools themselves, too.

In a single day I caught six silent defects in my own security scripts that were cheerfully reporting a green status on the outside. The leak scanner, run with the semantic-analysis flag, printed a header and exited with code zero without running a single detector. Elsewhere, an empty parameter silently disabled the scanner entirely, and the word "Anthropic" appearing in a line hid a live key sitting right next to it.

The fix is one trick: a check must be able to fail. Now, before every run, my scanner plants a control "canary" on itself with a synthetic secret and requires all 179 rules to catch their target. If even one stays silent, the entire report is rejected.

How do you take a ready-made architecture of 361 skills instead of

building it from scratch?

Current models already do far more than standard chat interfaces let them. When you strip away the excess product scaffolding and give the model direct access to a terminal, a browser under a working login, a slide builder, video editing, and a local catalog of 2,061 n8n scenarios, the speed of work changes by an order of magnitude.

I've published the entire rebuilt configuration — 361 skills and 39 agents — in the open on the [@jemal_hamidun](https://t.me/jemal_hamidun) channel. For those who want to deploy a stack like this locally with full privacy control, we're building out the [Nomad AI](https://nomadnet.ai) environment. You can work through orchestration practice step by step with a 72-hour trial at [Hamidun Academy](https://academy.hamidun.com), and regular masterminds plus the full library of 370+ skills are available at [Vibe Club](https://hamidun.com/club).