How to write instructions Claude actually follows

How to write instructions Claude actually follows

This is the most important skill you'll learn. The difference between Claude producing garbage and producing exactly what you want comes down to how you write your instructions.

The pattern that works every time

Every good instruction has these parts:

1. What to read — Tell it where the input is
2. What to do — Be specific about the transformation
3. What not to do — Explicitly tell Claude Code what to avoid
4. What to output — Tell it exactly what file to create

Here's the template:

Read [input file/folder].
[Specific instructions about what to do].
[Optionally: what not to do]
Save the result as [output file].

Bad vs. good instructions

Bad: "Summarize this data"

Good: "Read sales_q1.csv. Create a summary that includes: total revenue, average deal size, top 5 accounts by revenue, and month-over-month growth rate. Format as a markdown document. Save as q1_summary.md"

Bad: "Clean up this file"

Good: "Read contacts.csv. Remove rows where email is empty. Deduplicate by email address, keeping the most recent entry. Standardize all phone numbers to +1 (XXX) XXX-XXXX format. Save as contacts_clean.csv"

The specificity principle

Ask yourself: "If I gave this instruction to a new hire on their first day, would they know exactly what to do?"

If the answer is no, add more detail. Claude is capable, but it can't read your mind about your preferences.

Use examples when the format matters

If you care about the exact output format, show an example:

Format each entry like this:

**[Company Name]**
Contact: [First] [Last]
Revenue: $[amount]
Status: [active/churned]

Tell it what NOT to do

Sometimes it's helpful to set boundaries:

- Do NOT modify the original file
- Do NOT include rows where status is "test"
- Do NOT round the numbers

Break complex tasks into steps

For bigger jobs, number your steps:

1. Read all CSV files in the /data folder
2. Merge them into a single dataset, matching by email
3. For each person, calculate total spend across all files
4. Flag anyone who spent more than $10,000
5. Sort by total spend, highest first
6. Save as high_value_customers.csv

Next up

Now that you can write solid instructions, let's make them reusable. In the next episode, you'll create your first CLAUDE.md file — a saved instruction set you can run anytime.