Heidelberg AI Curriculum

T02-L01 · Chat & assistants · Level 1 User · 12-minute read, plus 25 minutes of doing

Longer, but still generic

You paste your notes into a chatbot and type "summarise this". The answer is fluent and could describe almost any project. You try again with a longer request and add "please be precise and professional". The second answer is longer and just as generic.

In a lab it's a postdoc with rough bench notes who needs a methods paragraph. The chatbot writes that the samples were measured "under standard conditions". Nobody wrote down any standard conditions, and the one threshold that matters has gone vague.

In a company it's an account manager with meeting notes who needs a client update. The chatbot writes that the project "is progressing well" and that the open checks "will be finished before delivery". There is no delivery date in the notes.

Neither of them needs a cleverer phrase. The chatbot had to guess what the notes were for, who would read the result and what mattered, so it guessed. This book shows you what to put in one message so it doesn't have to guess, and how to improve the answer with follow-ups that teach you something. You're at Level 1 User: if an answer is wrong, only you see it. That makes this the place to practise.

After this you can

Put the material, the reader, an example and the exclusions into one request.
Improve an answer with a follow-up in the same chat that changes one thing.
Say which change made the difference, and keep it for next time.
Pick the right kind of assistant for drafting, searching or analysing.

Before you start

Books: T01-L01. This book applies its four quadrants to a single conversation.
A chat assistant. If your lab or company provides one, use it. Otherwise use Claude on the free plan. It's good at careful drafting from material you hand it, and the free plan lets you attach files. ChatGPT's free plan is a fine alternative. Either way, check the training setting first (T01-L01, fact 4).
Half a page of your own rough notes that you're allowed to paste: meeting notes, bench notes, the points for an email.
A text file or note for your before/after record.
40 minutes: 12 to read, about 25 to do.

Last verified 23 September 2026. Claude's free plan runs in the browser and in the desktop and mobile apps, and includes the Sonnet and Haiku models. To attach a file, click + at the lower left of the message box and choose Add files or photos (up to 20 files per chat), or drag the file into the chat. ChatGPT's free plan allows unlimited everyday text chats, with separate, tighter limits on file uploads and data analysis. Perplexity's free plan puts citations in every answer. Gemini's free plan works with any Google account. If a button moves, the method stays the same: attach the material, then say who it's for.

The idea in one page

A chatbot's answer can only be as specific as your request. Four moves take the guessing out, in this order:

Move	Instead of	Write
1. Give the material. Paste or attach it. Don't describe it.	"Summarise my notes from Tuesday's meeting."	The notes themselves, marked off: `<notes> … </notes>`
2. Say who it's for and how long. Name the reader and what they'll do next.	"Make it clear."	"For a client who hasn't seen the project. Three bullets, under 60 words."
3. Show one example of what good looks like. A pattern beats ten adjectives.	"Make it professional."	"Use this form: Done: … / Open: … / Your decision: …"
4. Say what to leave out. Name the mistake you expect.	Nothing.	"Use only the notes. Add no dates, promises or recommendations. If a detail is missing, write not stated."

Move 4 is the one people skip, and it prevents the worst errors. A chatbot fills gaps with plausible text (T01-L01, fact 1). "Not stated" gives it something honest to write instead.

Then improve it, one change at a time. Stay in the same chat. The chatbot sees the earlier messages in the conversation, so a follow-up like "keep the facts, make it three bullets" works without repeating everything. Change one thing per follow-up. If you change the reader, the length and the format at once and the answer gets better, you won't know why, and you can't repeat it next week. Label the answers V1, V2, V3 and write one line under each: what you changed and what it did. Start a new chat only when the task or the material changes.

Use the right kind of tool. One assistant doesn't do every job equally well. For a first pick:

The job	Start with	Why
Draft, rewrite or summarise material you have	Claude	Careful with long texts, and sticks to "use only this"
Find current public information	Perplexity	Every answer links its sources
A quick look at a spreadsheet	ChatGPT	Can run calculations on an uploaded file
Work in Google Docs, Gmail or Drive	Gemini	Built into Google's apps

A search tool gives you sources, not finished text. A drafting tool gives you text, not current facts. T16-L01 goes deeper into choosing and testing a tool.

Worked example: rough notes to a finished text

Both versions take the same path through the quadrants. Follow the one closer to your work. The answers shown are typical examples written for this book.

Lab: bench notes to a methods paragraph

Structure it. "I need a methods paragraph for our internal report, written from my bench notes. The reader is a new lab member who may repeat the assay. A useful result is accurate and complete, and claims nothing the notes don't say."

Assay run, plate 3
4 samples, S-01 to S-04
mix 30 s
2 readings per sample
report mean; if the 2 readings differ by >5%, flag "review needed"
temperature: didn't note

Understand it. She pastes the notes with the request she'd normally type: "Write this up as a methods paragraph." V1 comes back:

Four samples were prepared and measured under standard conditions. Two readings were taken per sample and averaged to ensure reliable results. Samples with inconsistent readings were reviewed.

It reads well, and three things are wrong. "Standard conditions" is invented. The 5% rule has become "inconsistent". The missing temperature has vanished. The chatbot wasn't careless: she never said who the paragraph was for or that it had to stick to the notes. Seeing V1 changes her problem statement. The hard part was never the wording. It's keeping every number and every gap visible. Back to Structure it: "…and anything missing stays visibly missing."

Choose tools. A chat assistant is the right size. The material is half a page, she can check every line, and the notes hold no personal data. A search tool would be wrong, because nothing needs looking up. She stays in the same chat so she can compare versions.

Build it. Three follow-ups, one change each.

Revision 1, reader and length: "Rewrite this for a new lab member who may repeat the assay, 40 to 60 words." It now speaks to the right reader. "Standard conditions" is still there. A target sets the direction; it doesn't set a limit on the facts.

Revision 2, exclusions: "Keep that. Use only my notes. If a detail is missing, write not stated."

Mix each of the four samples (S-01 to S-04) for 30 seconds. Record two readings per sample and report the mean. If the two readings differ by more than 5%, mark the result "review needed". Temperature: not stated.

The invented condition is gone and the exact threshold is back.

Revision 3, example: "Keep that. Use this form: Preparation: … / Measurement: … / Exception: … / Missing: …" The facts stay the same, and she can now check the result line by line.

She checks each line against the notes, then asks a colleague for the temperature. No prompt could have supplied that fact. Her record says: Revision 2 fixed it. "Use only my notes, write not stated" removed the invented condition.

Company: call notes to a client update

Structure it. "The client wants a short update after today's call. A useful result is three accurate lines they can read in a minute, ending with the one decision they need to make."

Harborlight website project, call 12 Sept
done: draft navigation, 2 test pages
open: mobile review, accessibility check
client to pick 1 of 2 colour directions
delivery date: not agreed yet

Understand it. He types "Create a polished client update." V1 says the navigation is "complete", that the open checks "will be finished before delivery", and recommends "the stronger colour direction". Three problems: draft became complete, a delivery promise appeared, and the chatbot made the client's choice for them with no information about either option. "Polished" told it nothing about what it was allowed to say. Back to Structure it: "…and it promises nothing that isn't agreed."

Choose tools. Same as the lab: one chat, notes pasted in. The notes name a project but no people or prices, and he uses the assistant his company approved. On a personal account he'd replace the project name first.

Build it. Revision 1, reader and length: "Three bullets for the client, under 60 words." The length is right; the invented promise stays. Revision 2, exclusions: "Use only my notes. No dates, promises or recommendations that aren't in them. Write not agreed yet where that applies." The upgrade, the promise and the recommendation disappear. Revision 3, example: "Use this form: Done: … / Open: … / Your decision: …" The client's decision is now the last line, where they'll see it.

He checks each bullet against the notes and sends the update himself. The chatbot wrote text. It didn't send anything, and it didn't finish the open checks.

In both cases the first answer was worth having for what it showed: which instruction was missing. And in both, the change that mattered was move 4. The format made the result easier to use, but only the exclusion made it true.

What goes wrong

Describing the document instead of giving it. Symptom: "my notes from Tuesday" gets an answer about typical Tuesday meetings. Fix: paste or attach the text. The chatbot can't see a file you only mention.

"Summarise" with no reader. Symptom: the wrong length, the wrong vocabulary, the wrong details. Fix: name who will read it and what they'll do next.

Rewriting the whole request every time. Symptom: V4 is better than V1 and you can't say why. Fix: go back to the last answer you liked, change one thing, and write down what it did.

Searching with a drafting tool, or drafting with a search tool. Symptom: a chatbot's "current" figure that nobody can trace, or a search summary pasted into a report as finished text. Fix: find sources with a search tool and open them; draft from material you have with a drafting tool.

Accepting the first fluent answer. Symptom: a well-written paragraph with one invented condition leaves your desk. Fix: check every name, number and date against your material before you use it (T01-L02).

Believing it did something. Symptom: "I've sent the update" or "saved to your file". Fix: a chat assistant writes text. You copy it, save it and send it.

Do it yourself (25 minutes)

Structure it (4 min). Take your notes. Write one sentence: who will read the result, what they'll do with it, and what a useful result looks like. Copy the notes, unchanged, into your record file.
Understand it (6 min). Open a new chat. Paste the notes with the short request you'd normally type. Save the answer as V1. Mark every sentence the notes don't support and every gap that disappeared. If you learned what really matters, rewrite your sentence from step 1.
Choose tools (2 min). Check that you're in a drafting tool, not a search tool, and that you're allowed to paste these notes into it (T01-L01, fact 4).
Build it (13 min). In the same chat, send up to three follow-ups, one change each: reader and length, exclusions, example. Save the answers as V2, V3 and V4, with one line under each saying what changed. Stop when the answer is ready to use, and check it against the notes one last time.

Exit check

Delivered artifact: a before/after pair from your own notes: V1, your final version, and the single follow-up that fixed it, word for word.

It passes when a colleague can read the two versions side by side, see the improvement, and find nothing in the final version that isn't in your notes.

Rule to remember

Attach the thing. Then say who it is for.

Get Answers You Can Use from a Chatbot