Separating Instructions from Data & Not Assuming Shared Context

Providing Data Alongside Instructions

Many real prompts mix an instruction with the material to work on -- a transcript, a policy excerpt, a spreadsheet paste. When that happens, separate the two clearly. A heading like "Source document:" followed by the pasted text keeps Claude from confusing part of the data with part of the instruction. Without a clear boundary, the model may treat a sentence inside the source text as if it were a command from the user.

Claude Only Knows What You Tell It

Claude does not have implicit access to your organization's internal context -- its acronyms, unwritten policies, product names, or team structure. It only knows what appears in:

1. The prompt itself
2. Any Project knowledge you've attached
3. Its general training data

If a prompt assumes Claude already knows "our Q3 pricing exception process" without stating it, the output will fill the gap with a generic or incorrect guess rather than your actual policy.

Common exam traps

• Mixing instruction and data with no separation. A prompt that pastes a transcript directly after an instruction, with no heading or delimiter, risks the model reading part of the transcript as an instruction. Always label source material distinctly.
• Assuming Claude knows your organization's context. It only knows what's in the prompt, the Project knowledge, and its training data -- unstated internal context has to be supplied explicitly, every time it's relevant.