Using your own LLM

For plenty of everyday study tasks, your own preferred model is an excellent partner — no special setup from us required. Use whichever you like (ChatGPT, Claude, Gemini, Copilot, …); for anything with numbers in it, prefer one that can actually run code, because a model that computes an answer in a sandbox is far more trustworthy than one that merely predicts it.

These are good uses precisely because they do not need us to withhold anything. Where the value comes from holding back an answer — an exercise you are stuck on, a readiness check — reach for the Cogniti agents instead.

1 Good uses

The examples below show the difference a well-formed prompt makes. The pattern is always the same: give the model your material, name the specific thing you want, and — where you can — keep it from simply handing you the answer.

1.1 Explaining a concept you have read but not grasped

Try:

Here are my course notes on the chain ladder [paste them]. I follow how the development factors are calculated, but I don’t understand why we take cumulative-to-cumulative ratios rather than incremental ones. Walk me through just that step, using the notation in these notes.

Why it works: you have given the model the actual material, so it uses our definitions and notation, and you have named the precise point of confusion — so you get a targeted explanation instead of a generic lecture.

Avoid:

Explain the chain ladder method.

Too broad, and it hands the model nothing of ours to work from. It will fall back on its own textbook version (whose notation may differ from ours) and give you a recap you could have read in any reference — not help with your actual sticking point.

1.2 Generating extra practice

Try:

Here is a past CM2 question on ruin theory [paste it]. Write three new practice questions at the same level and style, changing the numbers and the claim-size distribution. Give me the questions only — hold the solutions until I ask.

Why it works: a real exam question anchors the difficulty and scope to the syllabus; asking it to withhold solutions keeps you working rather than reading; and because the papers are public with examiner reports, you have a genuine mark scheme to check its later answers against.

Avoid:

Give me some hard actuarial exam questions.

Nothing fixes the level or the topic, so you may get material that is off-syllabus or wrongly pitched — and you have no reliable way to tell, because there is no reference to check it against.

1.3 Debugging R or Excel formulas

Try:

This Excel formula should total the paid claims for accident year 2019 but it returns 0: =SUMIFS(Paid, Year, 2019) [paste the formula and a few rows of the data]. What is wrong, and how do I fix it?

Why it works: you have supplied the actual formula, a sample of the data, and the symptom (expected vs. actual). A model that can run code will reproduce the problem and confirm the fix, rather than guessing at what the formula would do.

Avoid:

Why isn’t my spreadsheet working?

No formula, no data, no error message — the model can only guess, and a numerical “fix” it hasn’t actually run is not something you should trust (see Watch out for below).

1.4 Critiquing your writing for a non-technical reader

Try:

I am advising a CFO with no actuarial background. Here is my recommendation [paste it]. Don’t rewrite it — instead, flag any sentence a CFO would not follow, and any jargon I should define or cut.

Why it works: it assigns a clear reader (the CFO) and a clear task (critique, not rewrite), so you keep authorship of the writing and learn what is unclear and why — which is the skill being tested.

Avoid:

Make this sound more professional.

Vague, and it invites the model to rewrite the whole thing for you: you learn nothing about your own clarity, and it may quietly introduce claims or caveats you never intended.

1.5 Quizzing yourself on a solution you have already worked through

Try:

I have just worked through this solution [paste it]. Ask me five short questions, one at a time, that test whether I actually understand each step. Wait for my answer before moving to the next one.

Why it works: one question at a time forces active recall, and basing it on a solution you have already completed keeps it as revision rather than a shortcut to new material.

Avoid:

Summarise this solution for me.

That is passive — you re-read rather than retrieve. It feels productive but builds very little recall, which is the opposite of what quizzing is for.

2 Watch out for

Notation drift. Models default to textbook conventions, which are not always ours. State our notation explicitly — the simplest fix is to paste the relevant notes.
Invented conventions. Models will state actuarial conventions confidently even when practice genuinely varies (reserving especially). Cross-check anything load-bearing against the notes.
Numbers. Never trust a numerical answer that was not computed by a tool. If it was not run in code, treat it as a guess.
Integrity. Nothing assessable goes in unless the assessment explicitly allows it, and the usual academic integrity rules apply.

A useful division of labour: your own LLM explains, drills and checks; the Cogniti Exercise Coach is the one to reach for when you are stuck on an exercise and don’t want to spoil the solution.