An analysis report on an interview with Boris Cherny, the creator of Claude Code, landed in my inbox in August. Core recommendation: delete over 80 percent of the system prompt; the model would work smarter without this baggage. It sounded plausible. So I tried it in my Claude Code configuration and disabled my 72 self-written skills.

The result was significantly worse.

What the report said

The report recommends a method Cherny calls "Ablation": completely delete the system prompt, let the agent work on the real project, log errors, and only then retrieve instructions if the same thing keeps going wrong. Along with numbers that are staggering. The Bun codebase, over 100,000 lines, rewritten from Zig to Rust in eleven days. Arc-AGI-3 from single-digit values to 30 percent.

I have retrieved the original transcript and compared it. At five points, the report flips the statement.

The "Lethal Trifecta" is listed there as a security architecture of three protective layers. In the interview, it is the opposite, namely the problem—the term comes from Simon Willison and describes a risk combination. The Electron-to-Swift migration is described in the report as "fully transformed" and ran "14 days completely autonomously." Cherny in the transcript: "It's still running." So not finished after two weeks. The Bun migration was allegedly "initiated by a single strategic prompt." To this exact question, he answers: "No, it wasn't one shot, but there was steering." From daily "20 or 30 of these routines," the report makes 20 to 30 pull requests. And the Max-20X quota, in the report a strategic lever for twentyfold performance, was a giveaway to the Startup School audience.

But the most important deviation is an omission. Cherny says about the deleted prompt:

"the model is actually a little bit more intelligent without these prompts. But when you use Claude Code as a product, you do actually want some of these prompts because it helps you use the product and it helps the model behave in the way that you would want."

The second half of the sentence is missing from the report. From a method with three steps, this becomes a deletion instruction with one.

Why my test performed worse

The difference lies in the type of instruction. Cherny deletes behavioral corrections: 'read before write', 'use absolute paths', 'batch your tool calls'. Things that a current model does correctly on its own and that now only consume attention.

My skills contain something else. They include knowledge that cannot possibly be embedded in training: the architecture of my chatbot module, the implementation of the AI disclosure under EU AI Act Article 50, the YAML schema of my report generator, the conventions for module manifests in my client projects. Out of 72 custom skills, 37 are tailored to Odoo, FastReport, or my infrastructure. No model can guess that.

If you are working with a grown configuration, this distinction is worth making before you delete anything: Does it contain an instruction on how the model should work, or does it contain knowledge that it cannot obtain anywhere else?

What I measured

Before the overhaul, I counted instead of estimated. In my configuration, there are around 16,000 tokens in context before I type the first word: 33,869 characters for skill descriptions, 8,705 characters for agent descriptions, 13,279 characters for base instructions, plus a hook that writes skills and agents in a second time.

Half of that comes from a single extension that brings 71 skills, 34 agents, and 19 hook entries into my setup. I only need them for specific projects.

What I built

A profile switch has been integrated into my SCCS sync tool.sccs profile off gsd moves the skills and agents of an extension into a parking area outside the Claude directory, removes their hook entries from the configuration, and sets the status bar to a replacement preset. sccs profile on gsd retrieves everything.

Nothing is deleted. If a file in the parking area collides with a same-named one, the process aborts instead of overwriting.

Two details cost me more time during construction than the switch itself. First, my own maintenance tool reinstalls the disabled extension on the next run until it learns to skip disabled profiles. Second, the sync function would have started pushing the parked remnants to the Git repository on the first run if I had filtered the exclusion list at the same spot. A test now confirms this.

What the switch actually brings

Around 5,000 tokens. Not half of my setup, as I initially hoped.

The 71 skills of the extension have very brief descriptions; the bulkier chunk is my own 72 with 26.058 characters. Something else is more noticeable than the tokens: Eight of the 19 hook entries are guards that start a Node process on every read, write, and shell call. Every time. Even when the extension has nothing to do that day.

How it proceeds

I will now work with a disabled profile and see what I lack. This is Cherny's third step, omitted by the report. Disable, measure, and retrieve what turns out to be necessary.

Three items remain on the list. My rule file contains instructions like 'read before writing', exactly the kind Cherny rightly removes. The hook that duplicates skills and agents costs about 2,000 tokens with no return. And seven of my skill files exceed 40 KB, the largest at 80 KB, which means roughly 20,000 tokens in a single call.

Whether this ultimately yields noticeably better answers or just a prettier number in the context window, I don't know yet. I'll let you know when I do.