Getting the Best out of AI

The difference between an AI assistant that wastes your afternoon and one that saves you a fortnight is almost entirely in how you ask. There are no magic words. What works is a habit of mind closer to writing a specification than to holding a conversation: say precisely what you want, say how you will know it is right, and check the result.

This page is the practical part. It is worth reading before your first serious session.

Say what "done" looks like

A vague instruction gets a vague interpretation, and the assistant acts on its interpretation immediately.

Instead of:

Tidy up the Darwin family.

Say:

For each child of family F14, if the person has a baptism date in a
note but no baptism fact, add a baptism fact with that date and place.
Change nothing else.

The second version can be judged. You can look at the result and say whether it was done. The first cannot, and neither can the assistant.

Ask for a report before you ask for a change

This is the single most valuable habit, and it costs one extra message. Split every substantial job into a look and a do.

List the children of family F14 that have a baptism date in a note but
no baptism fact. Show me the dates you found. Do not change anything.

Read the list. If it has misunderstood, you have lost nothing and you can correct it in one sentence. If it is right:

Good. Add those baptism facts for the first three only.

You now have a small, checkable change rather than a large, mysterious one.

Name things exactly

The assistant works with record identifiers — I42 for a person, F14 for a family, S12 for a source. Using them removes any doubt about who you mean, which matters when your tree contains four men named Charles Darwin. If you are not sure of an identifier, ask for it first:

What is the identifier for the Charles Darwin who married Emma Wedgwood?

Names are fine for questions. Identifiers are worth the trouble whenever something is going to be written.

Work in small batches

One family at a time, one source at a time, one kind of fact at a time. Small batches are cheaper, they stop before the assistant runs out of room to work in, and — most importantly — you can actually verify them. A hundred changes you have not checked are worse than none.

Pin it to the evidence

Left to itself, a model will fill a gap with something plausible. Say where the information is to come from, and say what to do when it is not there:

Using only the transcript on source S12, add a residence fact to each
person named in it. If the transcript does not give a place for
someone, leave it out and tell me which ones you skipped.

"Tell me what you skipped" is worth adding to almost any instruction. It turns silence into a report.

Refine rather than restart

If a reply is close but not right, correct it in the same conversation — the assistant still has the context and a single sentence is usually enough. Start a new conversation when you change subject, not when you hit a small problem.

Verify

Spot-check the records it touched — ask it to navigate you to one or two of them. Run the audit or a query over the area it worked on. If something is wrong, Undo on the Edit menu will reverse recent edits, but a saved copy from before the session is the real safety net.

Weak and strong requests

Weak	Strong
Fix the dates in my tree	Which people have a baptism date before their birth date? List them, don’t change anything
Add sources to these people	Add a citation to source S12 for the birth fact of each person in family F14
Sort out this census	From the attached census page, list the household members with ages and occupations. I’ll tell you which to add
Is this the right John Smith?	What do my records give for John Smith I42 that could confirm or rule out the John Smith baptised at Shrewsbury in 1802?

What it is good at

Questions that span the data. "Which of Josiah Wedgwood’s grandchildren died before their fathers?" — tedious by hand, quick here.
Consistency sweeps. Impossible ages, mothers too young, burials before deaths, children born after a father’s death.
Data entry from a document. Attach a transcript or an image, get a draft, approve the parts you want.
Background knowledge. What a particular record type contains, how a boundary changed, what an unfamiliar occupation or abbreviation meant. Check anything that matters — but as a starting point it is often excellent.
Getting around. "Take me to Emma Darwin’s family."

What it is not good at

Being an authority. It has no idea which of the things it tells you are true. Facts about your family must come from records, not from the model.
Charts and report layout. It cannot see or arrange them. Those remain yours.
Sweeping instructions across the whole database. Not because it refuses, but because you will not check the result — and unchecked bulk changes are how a good tree quietly becomes an unreliable one.