Claude Code Prompts & Examples: The System Behind Effective Requests

Copy-paste prompts are just the beginning. Learn the underlying system—CLAUDE.md, slash commands, custom skills, and verification patterns—that makes every prompt work better.

The best Claude Code prompts follow three rules: give context via CLAUDE.md files, use slash commands like /compact and /model to control behavior, and structure requests as specific outcomes rather than vague instructions. Copy-paste examples are just the starting point — the real skill is understanding the system behind them.

This guide teaches you both. You'll get practical prompts you can use immediately, plus the underlying system — CLAUDE.md, slash commands, custom skills, and verification patterns — that makes all prompts work better. All examples work with the latest Claude Code models: Opus 4.6, Sonnet 4.5, and Haiku 4.5.

The Three Principles of Effective Prompts

Before diving into examples, understand these three principles that transform how Claude Code responds:

Setting Up Your CLAUDE.md (Do This First)

Before using any prompts, create your CLAUDE.md file. This is the single most important thing you can do to improve Claude Code's responses.

Run /init in Claude Code. It generates a starter CLAUDE.md you can customize. Here's an example structure:

# My Project Context

## File Organization Preferences
- Keep original files, create organized copies
- Use YYYY-MM-DD date format in filenames
- Lowercase with hyphens for folder names

## Data Processing Rules
- Always create backups before modifying CSVs
- Use UTF-8 encoding for all text files
- Round numbers to 2 decimal places

## Verification Requirements
- Show me sample output before processing entire datasets
- List what you'll do before doing destructive operations
- Count records before and after transformations

Now every prompt you write benefits from this context without you having to repeat it. Keep it concise -- Claude can reliably follow around 150-200 instructions total, and the system prompt already uses about 50 of those.

Essential Slash Commands

These built-in commands change how Claude Code operates:

File Organization Prompts

Start with Plan Mode (Shift+Tab twice) for these—let Claude show you what it'll do before doing it.

Look at ~/Downloads and organize files into subfolders by type:
- Images (jpg, png, gif, webp)
- Documents (pdf, doc, docx, txt)
- Videos (mp4, mov, avi)
- Audio (mp3, wav, m4a)
- Archives (zip, rar, 7z)
- Other (everything else)

First, show me how many files of each type exist.
Then show me the first 5 files that would go into each folder.
Only after I approve, do the actual move.

Scan ~/Pictures for duplicate images. Check by content hash,
not just filename.

Create a report showing:
- Groups of duplicate files with their paths and sizes
- Total space that could be recovered
- Which file in each group has the oldest date (likely original)

Don't delete anything. Just show me the analysis so I can decide.

Organize photos in ~/Pictures into year/month folders:
~/Pictures/2024/01-January/
~/Pictures/2024/02-February/
etc.

Rules:
- Use EXIF date if available
- Fall back to file creation date if no EXIF
- Move (don't copy) to save space
- Skip files already in year/month folders

After organizing, tell me: how many photos per year, and
which years have the most photos.

Data Processing Prompts

For data work, verification is critical. Always ask Claude to show sample output first.

Clean ~/data/customers.csv:
1. Remove rows with duplicate email addresses (keep first)
2. Standardize phone numbers to (XXX) XXX-XXXX
3. Capitalize first letter of first/last names
4. Remove rows where email is empty

Before saving:
- Show me the first 10 rows of cleaned data
- Tell me how many rows were removed and why
- Show any phone numbers that couldn't be standardized

Then save as customers_cleaned.csv

Combine all CSV files in ~/data/monthly-reports/ into one file.

Requirements:
- All files have the same columns
- Add a 'source_file' column showing which file each row came from
- Add a 'row_number_in_source' column for traceability
- Sort by date column if it exists

Show me: total files found, total rows combined, any files
that had different columns (don't include those).

Save as combined_report.csv

From ~/data/sales.csv extract rows where:
- status = 'completed'
- amount > 1000
- date is within the last 30 days

Transform:
- Convert date to YYYY-MM-DD format
- Round amount to 2 decimal places
- Add a new column 'days_ago' showing how old each sale is

Verify by showing me:
- Count of matching rows
- Date range of extracted records
- Min/max/average of amounts

Save to high_value_sales.csv

Creating Skills (Reusable Prompts)

Here's where Claude Code becomes powerful: you can turn any prompt into a reusable skill.

Create a folder at .claude/skills/ in your project. Each skill gets its own directory with a SKILL.md file. Unlike the older .claude/commands/ system, skills support additional features like supporting files, frontmatter metadata, and lazy-loading (only the description loads initially, saving tokens):

# .claude/skills/clean-csv/SKILL.md
---
description: Clean and standardize CSV files
trigger: user asks to clean a CSV
---

Clean the CSV file specified by the user:

1. Remove duplicate rows based on the first column
2. Trim whitespace from all text fields
3. Standardize empty values to empty string (not "NULL" or "N/A")
4. Remove completely empty rows

Before saving, show:
- Row count before and after
- Sample of 5 rows
- Any issues found

Save with "_cleaned" suffix.

Skills activate in two ways: Claude can automatically detect when a skill is relevant based on its description, or you can invoke them manually. Claude reviews available skill descriptions and loads the full instructions only when needed.

Example: Weekly Report Skill

# .claude/skills/weekly-report/SKILL.md
---
description: Generate weekly metrics report from CSV data
trigger: user asks for weekly report
---

Generate my weekly report from the data in ~/data/metrics/:

1. Read all CSV files from this week (last 7 days by filename)
2. Calculate: total revenue, avg order value, top 5 products
3. Compare to previous week (files from 8-14 days ago)
4. Create summary with week-over-week changes

Output as:
- Console summary (key numbers)
- weekly_report.html (formatted for email)
- weekly_data.json (for archiving)

Include a "generated on [date]" footer.

Note: If you still have files in .claude/commands/, they continue to work. But new skills should use the .claude/skills/ format for the additional features.

Automation Prompts

These create scripts that run without Claude Code.

Create a shell script for daily backups:

What to backup:
- ~/Documents
- ~/Desktop
- ~/Projects

Where: /Volumes/External/backups/

Requirements:
- Name folders with date: backup-2024-01-15
- Keep only the last 7 backups (delete older automatically)
- Log everything to backup.log
- Show progress during backup
- Exit with error code if backup fails

Save as ~/scripts/daily_backup.sh and make executable.
Show me how to schedule it with cron.

Create a Python script that watches ~/Dropbox/incoming/ and
whenever a new PDF appears:

1. Wait 5 seconds (ensure file is fully uploaded)
2. Rename with today's date prefix: 2024-01-15-original-name.pdf
3. Move to ~/Documents/processed/
4. Send macOS notification "Processed: [filename]"
5. Log action to watcher.log

Run continuously until Ctrl+C. Handle errors gracefully
(skip problem files, log the error, continue watching).

Save as ~/scripts/watch_incoming.py

Read ~/data/weekly_metrics.csv and create an HTML email report:

Include:
- Summary table with this week's key metrics
- Week-over-week changes with color (green up, red down)
- Simple bar chart using CSS (no JavaScript required)
- Top 5 items by revenue

Style:
- Clean, professional design
- Works in email clients (inline CSS)
- Responsive (readable on mobile)

Save as weekly_report.html
Also show me the key numbers in the console.

Research and Analysis Prompts

Best used in Plan Mode—Claude explores thoroughly before summarizing.

Explore ~/projects/new-codebase/ and give me an orientation:

1. What does this project do? (Based on README, package.json, etc.)
2. Tech stack and main dependencies
3. Folder structure—what's in each main directory
4. Entry points—where does execution start
5. How to run it locally (based on scripts and config)
6. Key files I should read first to understand the architecture

Don't explain every file. Give me the 20% I need to understand
80% of how this works.

Read my meeting notes at ~/notes/team-sync.txt and extract:

1. Meeting context (date, attendees—infer from content)
2. Key decisions made (not discussions, actual decisions)
3. Action items formatted as:
   - [ ] Task description (@owner, due: date)
4. Open questions that need follow-up
5. Topics that were deferred to next meeting

Format as clean markdown I can paste into our project tracker.
If owner or date isn't clear, mark as "TBD".

Analyze customer feedback in ~/data/survey_responses.csv
(the feedback is in the 'response' column):

1. Categorize each response: positive, negative, neutral
2. Identify the top 5 themes across all feedback
3. For each theme, show:
   - Count of mentions
   - Sentiment breakdown
   - 2-3 representative quotes
4. Highlight any urgent issues (strong negative + specific)

Output:
- Executive summary (3-4 sentences)
- Detailed theme analysis
- Raw categorized data as feedback_analyzed.csv

Mode-Specific Workflow

Use the right mode for each type of prompt:

Example workflow:

1. Enter Plan Mode (Shift+Tab twice)
2. Ask Claude to explore your files: "What's in ~/Downloads and how would you organize it?"
3. Review Claude's analysis and proposed plan
4. Switch to Normal Mode (Shift+Tab)
5. Say "Execute the plan"
6. Approve each action as Claude works

Using @ File References

One of the most useful prompt techniques: use @ to reference files directly. Claude reads the file content before responding, which is far more effective than describing where code lives.

Compare @src/utils/old-parser.ts with @src/utils/new-parser.ts
and tell me what changed.

Look at @package.json and suggest which dependencies are outdated.

Keyboard Shortcuts Reference

Verification Patterns

The most important skill: teaching Claude to check its own work. Add these to your prompts:

• Count verification: "Count records before and after transformation"
• Sample output: "Show me the first 5 results before processing all"
• Checksum/hash: "Verify the file wasn't corrupted during transfer"
• Round-trip test: "After converting to JSON, convert back and compare"
• Spot check: "Manually verify 3 random items from the result"

Common Mistakes to Avoid

• Vague paths: "Organize my files" → "Organize files in ~/Downloads/"
• Missing verification: Add "show me what you'll do first" for destructive operations
• Ignoring context: Set up CLAUDE.md instead of repeating preferences in every prompt
• Skipping Plan Mode: For complex tasks, research first (Shift+Tab twice)
• Repeating yourself: Turn recurring prompts into skills

Related Guides

• What Is Claude Code? — understand the core concepts
• Claude Code Tutorial for Beginners — your first real project
• Claude Code for Marketers — marketing-specific systems
• Claude Code for Data Analysts — data processing workflows
• Claude Code for Non-Coders — no programming required