Ask the archive: answers from your team's files, with the source attached
Ask the archive
In this lesson you'll set up an AI assistant to answer questions from a folder of your team's own files, and to name the file and page behind every answer. There's no code. It takes about 10 minutes:
- Read (4 min): why answers drift away from your files, and four rules that keep them tied down.
- Question (2 min): pick the finding you'd put in front of a client.
- Question (2 min): decide where the files are allowed to go.
- Read (2 min): the prompt to keep, and what the rest of this chapter adds.
You'll leave with a short prompt you can paste into any AI project that reads your files.
The situation
You work at an agency. One client's shared folder holds three years of work in 140 files: campaign recaps, customer survey summaries, and two brand guides, version 4 from 2023 and version 6 from 2025. A new account lead joins on Monday and asks what the team already knows about this client. Reading everything would take about two days.
Most AI assistants let you create a project or workspace, add files to it, and ask questions about them. A legal team does the same with matter memos and a clause library. Done well, it turns a two-day read into an afternoon of asking and checking.
Why the answers drift
By default the assistant blends your files with everything it learned before, and it doesn't tell you which part came from where. That leads to three problems. It fills gaps with answers that sound right. It names a file that isn't in the folder. And it quotes the 2023 brand guide as if it were current, because both versions are sitting right there.
Four rules that fix most of it
Put these at the top of the project's instructions, or at the start of your question:
Answer only from the files in this project. For every claim, name the file and the page or section. If two files disagree, show both and say which is newer. If the files don't answer the question, say "not in the archive" and stop.
Then check before you share. For anything that goes to a client, open the cited page and confirm it says what the answer says. It takes about a minute per claim, and it's the whole reason to ask for the source.