3 mistakes from my first month with Claude Code

The 30-second check: how long is your CLAUDE.md

CLAUDE.md is the instruction file Claude Code reads at the start of every session. The Anthropic docs recommend keeping each such file under 200 lines, because a longer file takes more of the session's context, and it follows the instructions in it less.

Open a new session and paste:

Check how many lines each CLAUDE.md file loaded in this session has, including ~/.claude/CLAUDE.md and every file it imports with @. Write me a table: where the file is and how many lines it has. Count with wc -l, not an estimate, and flag every file over 200 lines.

Working in the terminal? You can also run it directly, from the project folder:

wc -l ~/.claude/CLAUDE.md CLAUDE.md

How to read the result

• •Every file under 200: you are fine. Skip to sentences 2 and 3.
• •One file or more over 200: that is my first mistake. Sentence 1 fixes it.
• •A short file that imports other files with @: count them too. Per the docs, an imported file loads together with the file that imports it, so it saves no space.

The three sentences, to copy

1. When the file is over 200 lines

Shorten CLAUDE.md to under two hundred lines, and keep only what you must know for every task.

2. Before a big task

Ask me questions until you can write a plan, and start only after I approve it.

3. When it writes "done"

Show me the test you ran, and what it printed.

Mistake 1: an instruction file that only grew

What I did: after every feature I asked it to document it in CLAUDE.md. The git history of my first app shows the file grew from 153 lines on July 20, 2025 to 1,248 lines a month later. No one read it except Claude Code, and it stopped listening too.

What to do: make a backup first, because it writes over the file itself. In my test, in a folder without git, it overwrote the original with no copy.

Before you change CLAUDE.md, save a copy of it as CLAUDE.md.bak in the same folder.

Then sentence 1:

Shorten CLAUDE.md to under two hundred lines, and keep only what you must know for every task.

What I got: in a demo folder with a long file, wc -l measured 1,420 lines before and 63 after. The working rules stayed. The history of every feature, and one rule that was written 30 times, came out or shrank to a single line.

Watch out: in its summary it wrote "from about 1,000 lines to about 65". That is an estimate. Check the real number with the check from part 1, again.

And what about what came out? For instructions that matter only for one part of the project, the docs recommend a file in .claude/rules/ that loads only when it works on those files.

Mistake 2: one huge request

What I did: I asked it to fix one thing in my app's chat. On August 3, 2025 a single commit changed 31 files. There is no way to review 31 files, so I didn't.

What to do: before a big task, write what you want, and add sentence 2 at the end:

Ask me questions until you can write a plan, and start only after I approve it.

What I got: in a demo folder for a studio website I asked for "workshop registration with payment" plus this sentence. It came back with 13 questions, about the payment provider, invoices, limited seats and a waiting list. 0 files changed until I answered.

Why it works: the Claude Code docs recommend exactly this when you don't know yet what you want: ask it to interview you until it could write the plan, and only then build.

Mistake 3: I believed "done"

What I did: on August 13, 2025 it told me "production ready" three times in one day. I believed it. On August 15, 2025 I was already fixing a bug in the same code.

What to do: when it writes that it is done, don't ask "does it work?". Ask for the proof:

Show me the test you ran, and what it printed.

What I got: in a demo folder I asked it to fix a VAT calculation. It wrote "I ran the test and it passes". After this sentence it showed the test and its output, pass 1, fail 0, and added on its own:

"Before the fix I didn't run the test. With 0.17 the result was 117, so it should have failed, but I calculated that and didn't see it happen." (translated from Hebrew)

That is the whole point. "Done" on its own doesn't tell you whether anything ran. The Claude Code docs also recommend asking for evidence: the command it ran and its output.

The rule I still keep

More than a year later, this is written in my instruction file: a task is not done without proof. So you don't have to ask every time, save it once. Paste this into a new session:

A task is not done without proof. At the end of every task, show me the command you ran and its output. If you ran nothing, say so plainly.

Save this to CLAUDE.md, for all projects.

How you know it worked

• •The file: the check from part 1 shows under 200 lines.
• •The questions: after sentence 2 it asks, and changes no files until you write "approved".
• •The proof: at the end of a task it shows a command and its output, or says it ran nothing.

Didn't work? In order:

1. Open a new session. The file loads at the start of a session, not in the middle.
2. Ask "where did you save the rule?". If it is in the project file and not in ~/.claude/CLAUDE.md, write "for all projects" again.
3. Still too long? Run sentence 1 again, and the check after it.

Sources

• •CLAUDE.md, the under-200-lines target, @ imports and .claude/rules/: code.claude.com/docs/en/memory
• •Asking it to interview you, and asking for evidence before "done": code.claude.com/docs/en/best-practices
• •The numbers about my mistakes (153, 1,248, 31 and the dates) were measured in the git history of my first app.
• •The examples are real Claude Code output from clean demo folders, on November 13, 2026.