Published: 2026-10-09
Summary

Anthropic cut 80% of Claude Code's system prompt: how to audit your own skills and CLAUDE.md with /doctor

Chapters / key moments (click to jump — plays here on the page)

Anthropic says it removed more than 80% of Claude Code's system prompt for its newest models without a measurable loss on its evaluations. Nate Herk turns that into a practical clean-up for your own setup: keep the context only you can supply, delete rules written for older models, load procedures only when needed, then let Claude Code audit itself with /doctor. He shows what the checkup flagged in his project and the audit prompt he used.

Source video

"Anthropic Engineers Just 10x'd Everyone's Claude Code" by Nate Herk — Watch on YouTube → (the title overstates it: nothing in the video is measured as ten times anything)

What Anthropic reportedly changed

He cites an Anthropic engineer's post on context engineering for Claude 5 generation models: the Claude Code team cut over 80% of the system prompt for Opus 5 and Fable 5 and 5.1 with no measurable loss on its own evals. We could not reach Anthropic's original post to check the wording, but aiweekly and other write-ups report the same figure, so read the original before relying on it. His summary of the idea:

Rules written for weaker models become restrictive. His example: a hard limit on code comments was replaced by an instruction to follow the style of the surrounding code.
Use the smallest context that gets the result you want. "Minimal" doesn't mean short: if the model needs a detail, keep it.
Define what done looks like: the outcome, why you want it, and the constraints it must respect, then leave room for the work.
Load on demand. He cites Anthropic's finding that loading MCP tool definitions only when needed cut token use by 85% in their example and raised MCP evaluation scores (49% to 74% on Opus 4, 79.5% to 88.1% on Opus 4.5). Those are Anthropic's reported numbers for those older models. He also cites a 2025 paper where accuracy fell as input grew, but says that is no prediction for trimming your own CLAUDE.md.

What to keep and what to cut in your own setup

Keep what the model can't infer: your voice, goals, audience, how the business works, and the why behind a task ("a student resource guide" changes the analogies and examples it picks).
Cut rules added after one old mistake that now sit in CLAUDE.md, a skill and a third copy somewhere else, and anything Claude can look up itself, such as a long folder listing.
Keep CLAUDE.md as a starting point: what the project is for, the gotchas and landmines, and pointers to deeper instructions. Move detailed procedures, such as a whole video-editing process, so they load only when you are doing that job.
His experiment: he ran the same job (turn an interview into a student resource guide) with his full setup and with a fresh one. The full setup made the prettier, branded document; the fresh one produced the better structure, with ideas organised and timestamps added. His fix is a skill that fixes the branding (header, font, footer, logos) and states the intent, but leaves the content structure to the model.
He also repeats a remark attributed to Boris Cherny, creator of Claude Code, that you should delete your skills every six months. We haven't checked that against a primary source.

Let /doctor do the audit

/doctor
/doctor prompt-audit [path]

He runs /doctor in a new session. Claude Code's command reference describes it as a setup checkup that diagnoses installation, settings, extension and CLAUDE.md problems and proposes fixes Claude applies after you confirm; /checkup is an alias. The same page lists a prompt-audit subcommand that audits your instructions, requiring Claude Code v2.1.283 or later. He doesn't use it in the video, so test it yourself. What his run reported:

Skills table: each skill with where it is installed, how many recorded uses since installation, and an estimate of the tokens its listing costs (the front matter used to find the skill, not the full body that loads when it runs). Example sizes: 111 and 103 tokens, both with zero uses. It flagged 18 older skills with no recorded use and spared ones created days earlier. He still reviews the list before disabling anything, especially skills he uses through another tool. It drew on 50 recent sessions plus lifetime counters.
Duplication: a project map repeating a routing map, plus repeated template, reference and archiving sections: about 2,250 characters of cuts, or roughly 563 tokens.
Where memory comes from: project CLAUDE.md, user-wide instructions, the auto-memory index and the local CLAUDE.md. It also found a CLAUDE.md from a separate video project folder that was loading into his sessions, around 2,000 tokens he didn't know he was paying for.
Broken skills: two with problems Claude couldn't see past: a wrong file name, and a formatting error in the front matter so the description never loaded.
It also checked installation, automatic updates, auto mode, permissions and hooks, and offered to fix everything at the end.

The audit prompt, and the cadence

After the checkup he asked Claude Code to audit one workflow, read-only:

Audit this repo and show me the smallest change that would make the path
from a raw YouTube idea to a filming-ready script faster. Use the systems
already here. Cite the exact files you would reuse and do not edit anything yet.

It mapped the steps from idea to research, outline, script and teleprompter file, and found a handoff gap: the outline skill could gather facts but didn't require saving evidence in a file the script skill checked, while the script skill had a hard rule to re-verify facts. Of 25 video folders with a script, only three had a sources.md. It proposed three edits across two skills, and he approved them. His cadence: monthly if you add skills often, quarterly otherwise, and every time you switch to a new model. Once you trust it, run the checkup as a local scheduled task in the Claude desktop app and approve or reject the recommendations when you wake up.

Gotchas & Caveats

Do not delete a skill just because it shows zero recorded uses; usage counted only covers what Claude Code recorded, not other tools.
The 80% figure is Anthropic's report about its own product's prompt for specific models. It does not say your CLAUDE.md is too long; he tested only one workflow.
The video includes a sponsor segment and a community pitch, which are not part of the method.
Review any fix /doctor proposes before accepting it; fixes are applied after you confirm.

Key Takeaways

After a model upgrade, re-test your skills and CLAUDE.md against a fresh setup and delete the rules the new model no longer needs.
Keep context only you can supply, plus the reason behind the task; load procedures only for the job that needs them.
Run /doctor to find unused skills, duplicated instructions, stray CLAUDE.md files and broken front matter.
Make handoffs between skills explicit: write evidence to a file the next step reads.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Commands & Code Shown

/doctor

Purpose: In the video, run in a fresh Claude Code session, it audits skills, memory files, plugins, settings, installation, auto-update and permission/hook settings, and offers to fix what it finds.

When to use: Every time you switch to a new model, and monthly or quarterly otherwise, per the speaker. Check the Claude Code docs for exactly what your version reports.

Configuration shown

Prompt the speaker used to audit a workflow

Audit this repo and show me the smallest change that would make the path from a raw YouTube idea to a filming-ready script faster. Use the systems already here. Cite the exact files you would reuse and do not edit anything yet.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Run /doctor and read the skills table
It lists each skill, where it is installed, how many recorded uses it has had, and the estimated tokens its listing costs (the front matter, not the full instructions). In his run it flagged 18 older skills with no recorded use while keeping ones created only days earlier.
Check for duplicated and hidden memory
It showed where memory comes from: project CLAUDE.md, user-wide instructions, the auto memory index and the local CLAUDE.md. It found a separate project's CLAUDE.md in a desktop folder loading about 2,000 extra tokens, and duplicated sections worth about 2,250 characters (roughly 563 tokens).
Fix broken skills
It caught two broken skills: one with the wrong file name and one with a front matter formatting problem, so Claude was not seeing the description.
Audit a workflow with the prompt above
It mapped idea, research, outline, script and teleprompter, found that the outline skill did not require saving evidence where the script skill would look, and that the script skill re-verified facts already checked. It counted 25 video folders with a script but only three with a sources file, and suggested three edits across two skills.
Keep brand details, drop rigid procedure
Rebuild a skill so it carries only the header, font, footer and logos, and leaves content structure to the model, then compare against the old output.

Gotchas

Review the unused-skills list before deleting anything. Usage counts only cover what Claude Code recorded, so a skill you use through another tool can look unused.
Minimal does not mean short: keep any detail the model needs, such as your voice, goals, audience and why you are asking. Remove repeated rules, lookups the model can do itself, and procedures that only apply to other jobs.
The headline claim that Anthropic cut over 80% of Claude Code's system prompt comes from an article by an Anthropic engineer as described in the video; read the original before relying on it.
Research cited in the video: a 2025 paper found performance drops of 13.9% to 85% as input length grew (older models), and Anthropic reported an 85% token reduction and eval gains (49% to 74% on Opus 4, 79.5% to 88.1% on Opus 4.5) from loading MCP tool definitions only when needed. These are controlled experiments, not a prediction for your CLAUDE.md.
His own test: a stripped-down setup produced a messier but better-structured guide than the fully customised one, so the two were combined.
A local scheduled task in the Claude desktop app could run these checks on a schedule, per the speaker; the video includes a sponsor segment for HyperAgent.