Prepare Your Sources for AI: ChatGPT, Claude and Gemini

This course contains the use of artificial intelligence.

In 2026, many teams give ChatGPT, Claude or Gemini a folder of real work files in a project or notebook, and colleagues or clients rely on the answers. An assistant can answer confidently and still be wrong because of those files: a new version that is approved but not yet in force, a table that lost its headings in a PDF conversion, or a transcript that gave a suggestion to the wrong speaker. This course shows you how to prepare those files, and how to test what the assistant does with them before anyone relies on it.

Prepare a Source Pack Before People Rely on It

• Choose the files each question needs, and decide which one governs.

• Preserve the meaning when scans, recordings and spreadsheets are converted.

• Brief the assistant on its sources, conflicts and gaps.

• Test the answers, and the evidence behind them.

The method is Choose. Preserve. Brief. Test. It does not need code, and it does not depend on one tool. You build four things you can reuse: a source inventory, checked extracts, a source brief and a short test sheet. They make your decisions visible to the assistant and to the colleague who maintains the files after you.

You follow one fictional case from start to finish. At Northfield Partners, a small consultancy, the support team's assistant drafts client replies from a folder of fourteen files. Julia Hart, the account lead, and Adam Reed, a consultant, prepare those files for questions from their client, Tom Becker. The folder holds useful evidence next to old versions, suggestions, duplicates and conversion problems. Not every untidy file is a problem, and not every readable file is suitable, so you learn which passages to check.

Section 1, Choose, starts with a confident wrong answer. Tom asks how fast a priority ticket will get a reply, and the assistant quotes a four-hour clause that starts next month and covers only the Premium plan, although his company is on Standard. You start from the questions and the dates they concern, match each file to a question, leave out duplicates and irrelevant material, and write down the questions that no file answers. You separate approval from the date a source applies, decide which source governs for this client on this date, and check who can see the workspace before anything confidential goes in. You then record status, effective dates, scope, owner and location in a source inventory, with open questions on their own list.

Section 2, Preserve, checks the passages that could change an answer against their originals. A scanned fee table continues onto a second page without its headings, so in the extract Premium prices look like Standard ones. A meeting transcript swaps the speakers and turns fifteen seats into fifty, although every sentence reads well. A spreadsheet and a slide deck show how filters, merged headings, units and speaker notes carry meaning. You choose a better export, keep a checked extract or record the gap, and you keep the originals for checking. You also see a crooked scan that converted perfectly, because often the conversion is fine.

Section 3, Brief and Test, writes the instructions apart from the evidence: the task, the sources, the rules for choosing, a conflict rule and a line for gaps. You write expected answers and locations for five types of question before you ask the assistant anything, and your test sheet keeps "reviewed on paper" apart from "tested in this tool on this date". An illustrative run shows how to trace a failure, repair the cause and test again. You then keep the pack current with an owner, a review date and a change log, keep a labelled older version for questions about earlier dates, and re-run the affected tests when a source, the tool or its model changes.

Section 4, Test, takes a smaller pack from the same case into ChatGPT Projects, Claude Projects and Gemini Notebook, with the same six questions and the same brief, plus one language line in Gemini Notebook. You see where the files and instructions go in each tool, then judge every answer and its evidence separately. Most answers were right. The problems were in the evidence: a location that a colleague could not easily use, a spreadsheet export whose rows Claude could not read until a PDF table was added, and a citation that did not support its claim. These runs were recorded on 7 October 2026 with the settings shown. No unprepared pack was tested, so the runs do not measure what preparation adds. They are not a ranking, and tools change, so the dated tool guides in the resources show the setup steps. Microsoft 365 Copilot is covered by a dated guide, not a demonstration.

Practice is built into the course. Practice 1 gives you a new folder of refund questions for an inventory and a gap list. Practice 2 gives you page pairs, an audio clip of a call and a spreadsheet export to check. The Final Practice Pack asks you to produce all four deliverables for a new folder and compare them with a separate key. Templates cover the inventory, the conversion checks, the brief, the test sheet and the change log. Four section quizzes and a final assessment of eleven new situations apply the method, and every answer is explained. The course has fifteen videos and about 54 minutes of video.

The preparation exercises in Sections 1 to 3 work on paper with the files provided, so you do not need an AI account to learn the method. Before you rely on an assistant at work, you also test it in a tool your organisation has approved. The course does not teach coding, RAG pipelines, chunking, embeddings or prompt tricks, and it does not promise error-free answers.

This course is for people who give an AI assistant work files that other people rely on, in support, operations, consulting, enablement, and learning and development. Bring one folder you know well, and the questions it should answer.

The tools will change. The method stays the same.