Lab 4 · intermediate · About seventy-five minutes

Create a source-backed research workflow

Collect the sources yourself, separate fact from inference, keep a citation log, and verify every claim that survives.

You finish with

A short researched piece where every retained claim traces to a source you personally opened.

Before you start

  • A question worth answering.
  • A spreadsheet or table.

Material: Publicly accessible sources. Respect paywalls and licensing.

  1. Write the question precisely

    "Is X good" is not researchable. "What does UK guidance currently recommend for X, and when was it last updated" is.

    Checkpoint Your question names a scope and a time frame.

  2. Gather sources yourself

    Find and open six to ten sources. Save the URL and the date you accessed it. Do not ask a model to supply citations at this stage.

    Checkpoint Every source is one you have actually opened.

  3. Build the citation log

    Columns: claim, source, URL, date checked, read it yourself (yes/no), confidence. The fifth column is the one that matters.

    Checkpoint The log exists and has a row per claim.

  4. Summarise from the sources only

    Supply the documents and instruct the model to answer only from them and to say plainly what they do not cover.

    Checkpoint The output names its gaps rather than filling them.

  5. Separate fact from inference

    Go through the draft and mark each sentence: stated in a source, inferred by you, or unsupported. Delete or label the third category.

    Checkpoint Every sentence is classified.

  6. Verify the survivors

    Check each retained claim against the source once more. Quotations especially — this is where paraphrase drifts.

    Checkpoint Every claim confirmed against the source text.

Source-bounded summary
Use only the documents I have attached. Do not use anything else you know.

For every claim, cite the document and quote the exact sentence supporting it.

Where the documents do not answer part of my question, write "not covered in the supplied sources" and move on. Do not fill the gap from general knowledge.

At the end, list what a complete answer would need that these sources do not contain.

Question: [yours]

What finished looks like

A citation log of ten to twenty rows and a short piece where every claim maps to a row. The list of gaps at the end is often the most useful output.

Test criteria

  • Every claim in the final text has a log row.
  • Every log row has a URL and a check date.
  • Every quotation appears verbatim in its source.
  • The gaps are stated rather than hidden.

If it goes wrong

A citation that does not exist.
The model was asked to recall rather than read. Supply sources; never accept generated citations.
A paraphrase that overstates the source.
Require quotations. Compare the quotation to the paraphrase.

Privacy and safety

If your sources include personal data, the log is now a record containing it. Store it accordingly.

Sources3 referenced
  1. On the dangers of stochastic parrots: can language models be too big?

    Bender, Gebru, McMillan-Major & Shmitchell, FAccT (2021) · peer-reviewed paper · link checked 2026-09-06

  2. Claude developer documentation

    Anthropic · official product documentation · link checked 2026-09-06

  3. AI Risk Management Framework

    NIST · standards framework · link checked 2026-09-06

Links are checked at build time. If one has moved, tell us.

Read alongside

Next lab: Turn a repeated task into a reusable prompt and context pack