AI prompts for mortgage brokers

The gap between a useless output and a genuinely helpful one is nearly always the instruction, not the model. Brokers who try one of these tools, get something generic back and conclude it is overhyped have usually typed a nine-word request and expected it to know how a UK intermediary works.

Writing better instructions is a learnable skill and it takes about an afternoon.

The five things a working instruction contains

Almost every good prompt has the same components, in some order.

Context

Who you are and what the situation is. Saying that you are a mortgage adviser at a directly authorised firm in the UK changes the output more than anything else you can type, because it stops the tool writing for a general consumer audience in an unspecified country.

The task, stated as an action

Rewrite, shorten, list, compare, draft, extract. Vague verbs produce vague results.

The material

The actual text you want it to work on. This is the part people leave out, and it is why so much output reads as though it were written about nothing in particular. If you want a good summary of a lender's criteria note, paste the note.

Constraints

Length, reading level, tone, what to avoid. "Under 120 words, plain English, no jargon, do not use the word kindly" is worth more than three sentences of encouragement.

Output shape

Say whether you want prose, a short list, a table of considerations described in words, or a single paragraph. Ask for the shape you will actually use.

What a real one looks like

Here is the difference in practice.

Weak: write an email chasing documents.

Better: I am a UK mortgage adviser. Draft a short email to a client who has not sent their last three months of bank statements for the current account their salary is paid into. The lender needs them to verify the source of the deposit. Explain that in one clause. Ask for the full PDF statements rather than a screenshot, and point them to our secure upload link. Under 120 words, warm but direct, plain English, no jargon, sign off from Sarah. Do not state any lender's requirements beyond the one I have given you.

The second version gets you something sendable. The first gets you a template you will rewrite entirely.

That final sentence is worth copying into most of your instructions. Telling the model not to add facts you have not supplied is the single most useful constraint for regulated work, because unprompted specifics about criteria, rates and timescales are where these tools cause real damage.

Instructions worth building once

Do not compose from scratch each time. Build a small set your firm keeps and reuses, tuned to your voice.

• A rewrite instruction that turns adviser shorthand into a clear client-facing explanation, at a specified reading level.
• A summarising instruction for long email threads that returns what was agreed, what is outstanding and who owes what.
• A structuring instruction that takes messy notes and organises them under your firm's own note headings.
• A challenge instruction that reads a piece of your writing and lists what a cautious compliance reviewer would query, without rewriting it.
• A translation instruction for turning a technical lender communication into something a client can act on.

Store them where the team can get at them and treat them as living documents. When someone finds a phrasing that reliably improves the output, it goes in the shared version rather than staying on their laptop.

What never goes into the box

Client names, addresses, dates of birth, account numbers, employer details, income figures, health information, credit history, and the contents of documents.

Almost every drafting task can be done with the identifying detail stripped out. The tool does not need to know the client is called Priya and earns £47,300 in order to write a good chaser about a missing payslip. Describe the situation in categories and put the specifics in yourself afterwards.

Where a task genuinely requires client data, it belongs in a tool your firm has assessed and contracted with as a processor under UK GDPR, listed on your record of systems, with retention understood. Not a personal account, and not a free tier whose terms permit inputs to be retained and reviewed.

Getting better output when it is not working

The instinct is to rephrase the whole request. Usually it is more effective to add one thing.

If the output is generic, you have not given it enough material. Paste more source.

If the tone is wrong, give it an example of your own writing and ask it to match that, which works far better than describing the tone in adjectives.

If it keeps producing something you do not want, say so explicitly as a rule rather than hoping it stops. Negative instructions are honoured reasonably well when they are specific.

If it is inventing details, tighten the boundary: tell it to use only the information provided and to write a placeholder where something is missing rather than filling the gap. Then look for the placeholders, which is a much faster review than hunting for fabrications.

If the answer is long and hedged, ask for the short version and then for the reasoning separately.

Where better prompting will not help

No instruction makes a model a reliable source of lender criteria, product availability or current rates. Those change constantly, models are trained on old text, and confident wrongness in this area is common. Check criteria at source. Every time, without exception.

Nor will phrasing make it a safe place to reason about suitability. You can ask it to list considerations a person might weigh. You cannot delegate the weighing, and the fluency of what comes back is not evidence that the thinking is sound.

The realistic ceiling is a fast, tireless drafting assistant with no judgement and no accountability. Instructed well, that is worth having. Instructed badly, it is a machine for generating things you have to rewrite.