Prune your instructions

TL;DR
- Anthropic cut over 80% of its own coding tool's instructions with no measurable loss.
- I cut my main instruction file from 598 lines to 116, and a new chat now starts on less than half the load.
- Keep only what the AI cannot work out for itself. Put the rules that cannot bend in settings, not in words.
This guide is for people who keep a standing file of instructions for an AI. In Claude Code it is called CLAUDE.md. Other tools have their own name for it: custom instructions, a project brief, a system prompt.
The file grows. Each bad day adds a line. Nobody takes lines out.
Then a new model comes along that follows every line, word for word. The file that helped the old model now holds the new one back.
What Anthropic cut
Anthropic did this to its own tool first. In a post on its blog, it wrote that it "removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations" (The new rules of context engineering, Thariq Shihipar, Anthropic, July 2026).
Its documentation gives a size to aim for: "Target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence" (How Claude remembers your project, as of November 2026).
A longer file is not a safer one.
What I cut
In November 2026 I did the same to my own setup. My websites and the software behind them live in one code repository, and Claude Code does most of the building. Here is what changed.
| What | Before | After |
|---|---|---|
| The main instruction file | 598 lines | 116 lines |
| The memory index, read at every start | 146 lines | 39 lines |
| Memory notes | 134 files | 30 files |
| What a new chat carries before I type | about 105,900 tokens | about 47,500 tokens |
The last row is from Claude Code's own /context command, counting only what loads before the conversation starts. The instructions and memory alone fell from 61,900 tokens to 6,200.
To check nothing was lost, I gave a fresh helper fourteen test requests after the cut. Each one tested a rule that matters: do not read the password file, do not send an email, ask before dropping data. It passed all fourteen, in two separate runs.
The test for each line
Anthropic's best practices page says the file "works best when it only holds what Claude couldn't work out from the code itself". That includes how to confirm work, places where the code misleads, decisions that are not written down anywhere else, and "anything you've had to correct twice already" (Best practices for Claude Code, as of November 2026).
So for each line, I asked one question. Could the AI work this out by looking?
A list of every folder and what is in it: it can look. Out.
The history of why a rule exists: it needs the rule and one line of reason, not the story. Out.
A decision I made that is written nowhere else: it cannot look. In.
Words guide, settings enforce
Some of my lines were rules that must never bend. Do not read the file of live passwords. Do not push to the live site without the checks.
Words are the wrong tool for those. Anthropic's page explains why: an instruction the model reads comes with "no guarantee that it'll follow your instructions". A setting is different: "Claude Code applies this every time, no matter what the model decided to do" (Best practices for Claude Code, as of November 2026).
Lydia Hallie, who works at Anthropic, gives a test for which tool to use, in a talk from November 2026. If the AI gets it wrong once, can you take it back? If you can, words are fine. If you cannot, make it a setting.
I moved my hard rules into settings and a small checking script. The file now only guides.
Load a rule when it is needed
Not every rule needs to be read at every start. A rule about the database only matters when the AI is working on the database.
Claude Code lets you put such rules in their own files, tied to a folder. They load when the AI opens a file in that folder. I now have ten of them. My main file has a short table that says which one to open for which kind of work.
OctAcademy tested four different CLAUDE.md files against the same six tasks, in a video from October 2026. It shows how much the file's shape changes the work.
Memory: say what it is for
If your tool keeps notes between conversations, prune those too. Mine had grown to 134 notes. Many repeated each other, and six contradicted each other.
Anthropic's engineering team wrote that "it helps to tell Claude what memory is for, but not how to use it" (The components of agent memory, November 2026). So my memory now holds what was learned along the way: a vendor's quirk, a tool's trap. My standing rules live in the instruction files.
Do it again with every new model
Claude Code can audit the file for you. Type /doctor prompt-audit and it lists lines written for older models, rules that contradict each other, and paths that no longer exist. Anthropic's memory page describes it (How Claude remembers your project, as of November 2026).
I have put it on my list for every quarter, and after each new model. The file will grow again. That is fine, as long as someone prunes it.
