All video notes
Video notes · Claude Code memory

Boris Cherny Just Told Us To Delete Our CLAUDE.md Files

Video by The AI Automators on YouTube · 16 min · uploaded July 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with The AI Automators. Watch on YouTube

In short

The AI Automators walk through a talk by Claude Code creator Boris Cherny and push back on parts of it. The headline advice is not really to delete CLAUDE.md for good but to strip prompts, skills and hooks every few months and see what the newer model does without them. The video's best addition is the line between instructions a model can outgrow and enforcement you must keep. It misses a second line that matters just as much: project facts the model can never infer.

The argument

The video's argument is that Cherny's advice is technically sound but shaped by his position. New frontier models, he says in the talk, can run for days with little scaffolding, and Anthropic removed most of Claude Code's own system prompt when Opus 5 shipped, so users should run the same experiment: remove prompting, then add back only what measurably helps. The AI Automators agree with the method and with his warning against over-specifying tasks. They object to the claim that Claude Code can no longer be prompt-injected, to treating multi-week autonomous runs as normal when he has an effectively unlimited token budget, and to ignoring projects where the model, not the product, is the bottleneck.

Key ideas, with timestamps

  1. 0:30

    Runs of days or weeks, with caveats

    Cherny says that with auto mode, newer models can work for days or weeks without goal commands or orchestration scripts. The presenter accepts that harness needs are shrinking but notes at 1:02 that the result depends on the use case and that Cherny is not paying for tokens. That framing runs through the whole video.

  2. 1:02

    The prompt-injection claim

    Cherny describes three layers: an aligned model, a prompt-injection classifier on all traffic, and the auto mode classifier that blocks dangerous commands. With all three, he says, the team can no longer demonstrate an injection. The presenter calls the implied claim of immunity ridiculous, and at 15:03 predicts it will not age well.

  3. 1:33

    80% of the system prompt deleted

    According to the talk, Anthropic removed over 80% of Claude Code's system prompt for Opus 5, because instructions tuned for one model rarely transfer to the next. At 2:01 he demonstrates the --system-prompt flag with a joke Klingon prompt, and at 2:32 shows the CLAUDE_CODE_SIMPLE=1 environment variable, which strips Claude Code's own system and tool prompting.

  4. 3:03

    Ablation, applied to your own setup

    Anthropic's process is to delete all prompting, then restore it piece by piece while measuring each line's effect, and to do the same for tools. For users, the takeaway at 3:33 is to clear CLAUDE.md, skills and hooks every six months and watch what the model does. The presenter adds the key caution: hooks that block destructive commands, deny rules and compliance checks are enforcement, not advice, and stay.

  5. 4:04

    Overhang, hobbling and its opposite

    Product overhang is when the model can do more than the product exposes, and hobbling is when the product gets in the way. Cherny's origin story is that early coding tools offered only autocomplete and read-only chat, so Claude Code gave the model a terminal and let it write whole files. At 5:33 the presenter adds the reverse case, where the product asks more of the model than it can deliver and needs more scaffolding, not less.

  6. 6:31

    Stop over-specifying

    Cherny says detailed step-by-step instructions now hurt more than they help. His replacement is to define the task, the guardrails and the exit criteria, then let it work. The presenter, a developer, admits he is prone to being too prescriptive.

  7. 7:33

    Bun from Zig to Rust in 11 days

    The talk's showcase is the Bun runtime, which Claude Code runs on, being rewritten from Zig to Rust over 11 days with a dynamic workflow. Cherny estimates a human team would need well over a year. At 8:31 the presenter stresses what made it work: human steering throughout and a strong existing test suite to verify against.

  8. 10:32

    What a two-week run really costs

    Cherny had a native desktop app rebuild running for 14 or 15 days, driven through Claude in Slack. The presenter cites an experiment claiming a $200-a-month Max plan can yield around $8,000 of API-equivalent usage. At API prices, he argues, similar multi-week runs could cost an organisation tens of thousands of dollars or more, and small drift can compound by day nine or ten.

What it gets right

Instructions versus enforcement is the most useful distinction in the video. A line in CLAUDE.md telling the model to write tests is a nudge a stronger model may no longer need. A hook that blocks rm -rf or a deny rule on production credentials is a control. Ablation applies to the first kind only, and the video says so clearly.

Verification is the real multiplier. Both Cherny and the presenter land on the same point from different angles: long runs work when the agent can check its own work, as with Bun's test suite. Without that, more autonomy just means more drift.

The scepticism is proportionate. The presenter credits the method and the product insight but flags the claims that depend on Anthropic's token budget or on marketing. Viewers get the talk's ideas plus a sense of which ones transfer to a normal team.

What it leaves out

CLAUDE.md does two jobs, and only one is outgrown. Much of a good CLAUDE.md is not behaviour correction at all. It is project knowledge: the build and test commands, which folder is legacy, why a library was rejected, the deploy steps, who owns what. No model upgrade lets the model infer those. If you delete the file and the agent starts asking or guessing, you have learned that the facts were load-bearing, not that the model got worse.

There is no method for measuring. Anthropic can ablate against internal evaluations. Most users cannot. A workable version: pick three to five real tasks you run often, run each with and without a section of CLAUDE.md, and keep the section only where the output gets clearly better or faster. Delete style advice first; keep facts and enforcement.

The title oversells. Neither Cherny's quoted advice nor the video's own summary says to delete the file permanently. Both describe a periodic reset and re-test, which is a more modest and more defensible habit.

How this applies to memory across AI tools

Splitting steering from knowledge also suggests where each belongs. Steering prompts are model-specific and should be short and disposable. Project and domain knowledge is durable and is useful to every tool you use, not only Claude Code. Keeping it in one tool's instruction file means re-creating it for Cursor, ChatGPT or the next agent. A persistent store that any MCP client can query lets CLAUDE.md shrink to pointers and rules. See Claude Code memory tools for options.

For the research side of that store, BrainTube saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it instead of each keeping its own copy. The broader case is in persistent memory vs the context window.

Who should watch it

Watch it if you maintain a long CLAUDE.md, a pile of skills or custom hooks and want a reason and a rule for pruning them; 1:33 to 4:04 covers the system-prompt cut, the two flags and the ablation advice. The overhang discussion at 4:04 to 6:00 and the cost reality check at 10:32 to 12:00 are the other strong parts. Skip it if you have not used Claude Code yet, since it assumes you know auto mode, skills, hooks and dynamic workflows, or if you want the primary source: this is commentary on Cherny's talk, not the talk.

Frequently asked

Should I delete my CLAUDE.md file?
Not outright. Periodically remove the behavioural instructions and test whether a newer model still needs them. Keep project facts such as build commands, conventions and past decisions, and keep anything that enforces safety or compliance. Those are not things a better model outgrows.
What is prompt ablation in Claude Code?
Ablation means removing prompting, tools or configuration and adding pieces back one at a time while measuring the effect of each. Anthropic uses it on Claude Code's own system prompt when new models ship. Users can apply the same idea to CLAUDE.md, skills and hooks with a few repeatable test tasks.
What does CLAUDE_CODE_SIMPLE=1 do?
As shown in the video, setting this environment variable strips back Claude Code's built-in system and tool prompting. It is described as an internal ablation switch for checking whether that prompting earns its place. Behaviour may change between Claude Code versions, so check current docs before relying on it.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Go deeper

More video notes

Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.