Prompting for plain language: give AI the right grounding

Asking AI to write in plain language leaves a great deal unspecified. The model needs to know who will read the explanation, what they need to do and which facts must be preserved. Without that information, it can produce a fluent draft that does not fit the situation. It may also introduce a promise or condition that the source never contained.

In our AI writing training, we distinguish grounding from prompting. Grounding supplies the approved information and constraints that give the task its context. The prompt tells the model what to do with that material for a particular communication. Keeping these elements clear helps the team maintain the sources and review the instructions when something changes.

Establish the facts before asking for a draft

For financial content, the starting point is the information the organisation can accurately give the reader. That may include eligibility conditions, charges, deadlines and the consequences of taking no action. These facts need to come from approved sources, with a record of the version being used. Otherwise, a carefully written prompt may simply produce a polished explanation of outdated information.

The brief also needs to make gaps visible. If a product document does not explain a particular condition, the model should not fill the gap with a plausible guess. An instruction to use only the supplied facts and list missing information gives the reviewer something to check. It does not remove the need to compare the output with the source and resolve the outstanding question.

Describe the reader using what you have learned

Audience grounding explains the circumstances in which somebody will encounter the communication and the knowledge the explanation can reasonably assume. This needs to come from research and relevant customer information. A broad description such as “older customers” tells the model very little about the task or the support people need. It can also encourage assumptions that the team has no evidence to make.

For a renewal letter, the useful information concerns what customers already know about the cover and which changes they need to consider. The team also needs to understand any relevant language, accessibility or support needs. Those details help the writer decide what to explain and how. They should be supplied only where their use is authorised, with uncertainty recorded rather than turned into a customer label.

Define the structure and the language you need

Once the facts and audience are clear, decide how the explanation should be organised. A reader may need the main change first, followed by its effect and the available options. The detail can then support that initial explanation without burying it. Giving the model this order makes the task more specific than asking for a clear or friendly message.

The same applies to terminology and tone. Supply approved terms, their meanings and any situations in which a different expression is required. Explain the relevant writing rules, supported by examples the team has reviewed. This reduces the number of choices left unspecified and helps reviewers distinguish an acceptable variation from a change that affects meaning.

Use prompt blocks to make the instructions reusable

A reusable prompt works best when it separates information that stays the same from information that changes with the communication. The shared guidance may contain terminology and writing rules, while the task specifies the audience, product facts and required output. The prompt can then draw on that material without restating the whole content policy. This also makes it easier to see which instruction needs updating when the product changes.

Our training uses prompt blocks for the role, task, audience, journey, mandatory content, information order and checks. Each block answers a practical question about the work the model is being asked to do. Examples can demonstrate the desired pattern, provided they are relevant and do not contradict the current facts. The team still needs to test the combined instructions on the kinds of communication it expects to produce.

Review the output against the brief

The first review needs to establish whether the draft contains the required facts and preserves their meaning. Only then can the team judge how well the explanation is written. A model can help prepare a comparison, but its account of what it included is another output to verify. The approved source remains the basis for checking the content.

After resolving factual and wording problems, testing with intended readers can investigate whether the communication works in practice. People need to be able to find the relevant information, understand it and use it for the decision they face. The findings may point to changes in the draft, the template or the grounding itself. Our guide to testing financial documents with customers explains this part of the process.

Maintain the information that future drafts depend on

Reusable instructions create continuing responsibilities. Someone needs to own the source information, update it when a condition changes and identify which patterns may be affected. Monitoring can also reveal a repeated misunderstanding that needs attention in the shared guidance. Without that work, automation can repeat an old problem across many new communications.

We help content, legal and compliance teams connect their writing guidance with practical AI workflows. Explore our AI writing training or contact Narratology to discuss the communications your team needs to produce.

Further reading and collaboration

The four principles in our guide to plain language and ISO 24495 explain the wider framework. Our work on purposeful prompting includes collaboration with Firehead.