Separate Data from Instructions

Most real prompts contain two different things: your instruction (“rewrite this email politely”) and the data it applies to (the email itself). If the two are not clearly separated, Claude can mistake one for the other — with comic or costly results.

Prompt templates: same task, changing data

In daily work you often repeat one task on different inputs — summarize this report, then that one; rewrite this email, then another. The professional habit is to keep a fixed instruction (a “template”) and drop the changing data into it. The simplest example:

What goes wrong without clear separation

Now a realistic case. The instruction and the data are pasted together with nothing marking where the email starts:

The fix: wrap the data in markers

The solution is to wrap the data in simple angle-bracket markers — an opening marker like <email> before it and a closing </email> after it. Claude was specifically trained to recognize these markers as boundaries, and you can name them anything sensible: <email>, <report>, <data>.

One more trap: lists that blend together

The same confusion happens with lists. Suppose your instruction contains a remark — “Each is about an animal, like rabbits.” — followed by the pasted list. Written with a dash, the remark looks exactly like a list item, so Claude counts it as one and reports the wrong “second item”. Wrap the pasted list in <sentences> … </sentences> markers and Claude counts correctly every time.

Small details matter

A closing lesson from this chapter: Claude is sensitive to small details. Typos, stray dashes, and sloppy grammar genuinely make it more likely to misread you — it tends to mirror the quality of what it is given. Scrub your prompts the way you would scrub a formal memo.