Prompt Engineering vs Context Engineering

Prompt engineering vs context engineering

This is the single most important conceptual lesson in the course.

The distinction

• Prompt engineering is how you ask. The wording of the request, the role you assign, the format you want back.
• Context engineering is what's already in the room when you ask. The files Claude has read, the project conventions in CLAUDE.md, the tool results currently in context, the subagents available.

Both matter. But context engineering produces the 10× quality jumps senior engineers report.

Side-by-side

Naive prompting:

"Add input validation to the user signup endpoint."

Result: Claude makes assumptions about your stack, your validation library, your error format. 50% chance the result needs rework.

Engineered context:

(Context: project CLAUDE.md says "use Form Request classes for validation; sibling requests show the array-syntax style." Claude has already Read the existing LoginRequest.php for reference.)

"Add input validation to the user signup endpoint, matching the pattern in LoginRequest."

Result: First-try-correct code that matches your codebase exactly.

What good context engineering looks like

1. A short, sharp CLAUDE.md that captures conventions Claude couldn't infer from one file (covered in 4 lessons from now).
2. Specific file references in your prompt: "follow the pattern in X.php" beats "use our standard pattern."
3. Negative space — telling Claude what not to do is often more valuable than telling it what to do.
4. Loaded relevant files — if you know the change touches three files, mention them.
5. Removed irrelevant context — /compact between unrelated tasks.

The mental model

Think of Claude's context window as a desk. You can stack everything on it ("just throw the whole repo in there"), or you can lay out exactly what's needed for the task. The second produces better work in less time.

Try it

Take a recent task you ran with Claude that needed two prompts to get right. Re-run it once with engineered context: pre-Read the relevant file, mention the pattern by name, state one thing not to do. Compare results.