Writing the Cited-Answer Prompt

Number the retrieved chunks and instruct the model to answer only from them, tagging each claim with the source number that backs it.

Part of: Personal Notes Brain

You now have the top few relevant chunks. The next job is turning them into a prompt that makes the model answer only from those chunks and mark which chunk backed each part of the answer. That is the difference between a chatbot and a research assistant with footnotes. Numbering the sources Give every retrieved chunk a number, then reference that number in the system prompt's rules. The number becomes the citation marker the model is asked to use: [1], [2], and so on. The system prompt does the grounding Notice the shape. The system prompt carries the rule , which is cite every claim and refuse to guess. The user turn carries the data , which is the numbered sources plus the actual question. That separation is the same discipline as every other product in this track: rules and input never mix. Why citation numbers, not source names Numbers are compact and unambiguous. "According to [2]" is one token. "According to your March 4th standup notes" is many, and the model paraphrases it inconsistently from one call to the next. You map the number back to the real title and file path yourself after the model replies. That is the next lesson's job. Why this matters Without an explicit cit

Challenge: Number the Sources