From material to insight

Today’s finish line

Use an agent for one genuine research task and verify at least one important part of its output against the source.

The point is neither blind trust nor ritual scepticism. The point is a workflow in which fast work can still be checked.

Morning — references, and your first skill

You have already met plain text and git: .bib is a bibliography as text, .qmd mixes prose, code, and citations in one readable file, and .gitignore tells git which files to leave alone. We put all three to work — and package the result so you never explain it twice.

Gather what you need

Point the agent at a topic, a rough reading list, or a pile of half-remembered titles, and have it:

1. resolve each work to a real paper — title, authors, year, DOI (inventing a citation is failure, not a shortcut);
2. download the open-access PDFs into a gitignored references/pdf/ folder;
3. write a one- or two-line note on each into references/notes.md;
4. add a verified BibTeX entry to references.bib;
5. list what it could not get — with DOIs — in references/to-fetch.md, so you can grab those through the library yourself.

Then say: "turn that workflow into a skill." The agent writes the skill file; you never learn its format. Next time, "gather refs on X" just works.

Skills aren't essential — you can do most of this course without them. They earn their keep only for routines you repeat and want done identically.

Memory is the AI remembering you. A skill is you writing it down — so it happens the same way every time, and you can hand it to a colleague.

Clean what you have

Now the other direction. We start from a reference list with deliberate traps — nonexistent papers, wrong DOIs, missing details, inconsistent formatting — and ask the agent to identify every work, verify it against an independent source, produce clean BibTeX, and record what it could not confirm. Then run it on a bibliography from your own work.

A polished bibliography is not evidence of a correct bibliography.

Full walkthrough: handouts/references.md.

Afternoon — a literature sweep

Give the agent a folder of papers and a concrete comparison question. Begin with one paper, settle on useful fields, and only then scale up.

Possible columns:

• research question;
• setting and sample;
• method or material;
• central finding;
• limitation noted by the authors;
• page supporting the summary.

Ask which papers genuinely disagree and which merely ask different questions. Export the result as Markdown or CSV so it remains useful outside the session.

The honesty checkpoint

Choose one row that matters. Open the original paper and verify it manually. Check the page reference, wording, and strength of the claim.

If the row is wrong, that is useful course material. Diagnose why:

• Was the question vague?
• Was the PDF read incorrectly?
• Did the summary overstate cautious language?
• Did the agent infer something the authors never claimed?

Improve the workflow, not just the sentence.

Open lab

Bring one research mess:

• an untidy bibliography;
• a folder of PDFs;
• rough project notes;
• interview or meeting material you are allowed to process;
• a table that needs documenting;
• a draft that needs structural criticism.

Before sharing material with any service, stop and consider confidentiality, personal data, contracts, and research ethics.

Before you leave

Save one useful artifact and a short verification note. Tomorrow we use the same workflow to build teaching material—and make it presentable.