Last updated
Claude Prompting Guide
A practical reference for crafting effective prompts with Claude (Anthropic). Focused on sysadmin, IT ops, and scripting use cases.
Overview
Claude excels at sustained reasoning, code generation, structured output, and following multi-step instructions precisely. Key differentiators: large context window (up to 200K tokens), strong XML/markdown handling, and minimal tendency to hallucinate compared to smaller models.
Prompt Structure
Every Claude request has three components:
| Role | Purpose | Example |
|---|---|---|
| System | Sets behavior, persona, constraints | You are a senior Linux sysadmin… |
| User | The actual request or question | Write a cron job that rotates logs… |
| Assistant | Prefill to steer format | “`bash\n#!/bin/bash |
XML Tag Technique
Claude understands XML tags natively. Use them to separate context from instructions:
<context>
Server: Ubuntu 22.04 LTS, 16GB RAM, running Nginx 1.24 and PostgreSQL 15.
Issue: DB queries slowing down after 72h uptime: CPU stays under 40%.
</context>
<task>
Diagnose the likely cause and provide a step-by-step investigation plan
using only CLI tools available on the server.
</task>Sysadmin Prompt Templates
| Task | Prompt Pattern |
|---|---|
| Bash script | Write a bash script that [task]. Include error handling, set -euo pipefail, and comment each section. Target: Ubuntu 22.04. |
| Log analysis | Analyze this /var/log/syslog excerpt and identify: (1) root cause, (2) affected services, (3) remediation steps. <log>[paste]</log> |
| Ansible playbook | Write an Ansible playbook to [task] for hosts in group [group]. Use best practices: become: yes, tags, handlers for service restarts. |
| Cron expression | Generate a cron expression for: [schedule description]. Explain each field. Include the line ready to paste into crontab. |
| Nginx config | Generate an Nginx server block for [domain] that: [requirements]. Include security headers, HTTPS redirect, and rate limiting. |
Chain-of-Thought Prompting
Add Think step by step before answering. or Before writing code, outline your approach. to get more reliable, auditable answers for complex tasks.
Best Practices
- Be explicit about output format: “Return ONLY valid JSON, no prose.”
- State constraints upfront: OS version, package versions, no root access, etc.
- Use role priming: “You are a security-focused Linux engineer…”
- Break large tasks: ask for an outline first, then expand each section.
- For code review: paste the code inside <code> tags and ask for specific feedback.
- Use extended thinking (claude.ai) for architecture decisions or complex debugging.
Last updated October 2026.
FAQ
How do I write better prompts for this AI?
Give context, a clear task and the output format. XML-style tags help separate instructions from data.
Can it read long documents?
Yes. It handles long inputs well; put the document first and your question at the end.