Claude Code Prompting Framework
Use when writing prompts for Claude Code or agentic tasks to raise first-pass output quality. Provides a 4-part prompt structure (role, context, negative, verification) plus a self-verification pattern.
Drop the file into ~/.claude/skills/learned-claude-code-prompting-framework/SKILL.md and your agent knows it in every session.
Claude Code Prompting Framework (Role + Context + Negative + Verify)
When crafting a prompt for Claude Code or any agent task, always include these four levers. From Nate Herk's 'Claude Code for Normal People' course.
1. Role
Tell the agent who it is and what it does.
- Example: "You are a master content strategist and data analyst."
2. Context & Background
Give both general background AND task-specific context. Generic prompts get generic output — the private, non-public context (your business, priorities, metrics, avatar, history) is what makes output unique.
- Weak: "Help me write this email to my boss."
- Strong: "Help me write this email to my boss. Tread lightly — I've gotten in trouble twice this month. I'm asking for more time off. He's understanding but I feel guilty asking, so keep the tone respectful and light."
- For work tasks, drop in real files/data: product details, calendar, copy that worked, copy that flopped.
- Analogy: treat the AI like a summer intern — onboard it before expecting good work.
3. Negative Prompt (what NOT to do)
Explicitly state guardrails. Like a curious student, the agent will try things unless told not to.
- Example: "Do NOT invent statistics. Do NOT change the tone to sound corporate. Do NOT touch files outside the /brand-assets folder."
4. Verification — MAKE THE AI PROVE ITS WORK
The biggest quality lever. Ask yourself: "If a human handed me this, what would I do to approve it?" Then instruct Claude to do exactly that.
- Build a website form? "Open the site and test the form submission 100 times, including bad emails without an @domain. Fix any bug you find, then re-test. Report what you ran and why you're confident."
- Data task? "Take screenshots to verify the numbers match the source before finalizing."
- Self-verification raises first-pass quality from ~60% to ~80%, cutting the number of manual feedback loops you have to run.
Choosing agent vs simple workflow (before you even prompt)
- Deterministic if/then task (same input → same output, e.g. 'every 9am pull Stripe revenue, post to Slack') = simple workflow/script. Cheap, reliable, no AI needed.
- Messy input needing reasoning/generation (e.g. 'read customer emails, understand intent, draft tailored replies') = use an AI agent.
- Default to the simplest thing that gets the job done.
Cost awareness
Output tokens cost ~5x input tokens. Use cheaper models (Haiku) for fast/simple jobs, balanced (Sonnet) for most, expensive (Opus) only when deep reasoning is required.