Using large language models (LLMs)

Large language models (LLMs), such as ChatGPT or Copilot, can support content designers and anyone who writes public-facing content. They can improve clarity, consistency and structure, but they are not a substitute for content design expertise.

This guidance shows you how to use LLMs responsibly and effectively. LLMs can make whatever you're writing easier to read for the public, more quickly than if you were drafting and editing yourself. They can also help you better meet Scottish Government Content Standards, including the Scottish Government Digital style guide.

This guidance is aimed at showing you how to use LLMs to get the best results. LLMs are not perfect, so if you expect them to do all the work, what you'll get will likely include mistakes or be of poor quality.

When using an LLM, you must give clear prompts, check everything carefully, and apply human judgment. Using an LLM is not an automated 'one and done' process. It involves iteration, review and accountability.

You are responsible for the accuracy and quality of anything you publish, even if an LLM helped produce it.

What a large language model is

A large language model (LLM) is an AI tool trained to generate and understand text. It can summarise information, rewrite content or suggest clearer ways to explain something.

You do not need specialist AI skills to use an LLM. You interact with an LLM by writing instructions called prompts. The model follows these prompts to create, edit or summarise content. You do this through a simple chat interface.

When to use an LLM

You might use an LLM to:

• draft something for the public
• check if your draft meets government readability and content standards
• improve clarity and structure, by asking the LLM to rewrite your draft
• check content against the Scottish Government Digital style guide
• summarise long or complex information (although the result can be inaccurate)
• generate examples or alternative versions of content

Do not use an LLM for:

• personal, sensitive or confidential information
• legally complex or policy-critical content
• situations where you cannot check the accuracy of the result

Before you start

Make sure you:

• do not enter personal, sensitive or unpublished information
• understand that LLMs can make mistakes or 'hallucinate' (invent false facts), so check any changes before publishing
• read the Scottish Government Digital style guide so you can judge quality
• check your organisation’s policy on using AI tools

How to write an effective prompt

A prompt is a set of instructions you give the LLM. Clear prompts give better results. Prompt writing is a skill that improves with practice.

Prompts are not one-sentence commands. They are part of an ongoing collaboration between you and the LLM. A good prompt sets the context, tone, and expectations for the response. It can include examples, constraints, or a role for the model to follow.

Some tools have behind-the-scenes settings that influence how the LLM responds, but you do not need to worry about these. Focus on writing clear, specific prompts to guide the LLM.

Example of a simple starting prompt

When writing a prompt, be specific about:

• who the content is for
• what the user needs to do
• the tone and style you want
• any limits such as word count, format or platform

Think about the information, examples or roles you can include to guide the model. The more precise and detailed your instructions, the more useful the output will be.

Iterating and improving your prompts

LLMs rarely get it right on the first try.

You may need to refine your starting prompt to improve clarity, structure or style. Iteration is a normal and important part of working with LLMs.

When refining a prompt, you might:

• ask for a shorter or simpler version
• ask for a clearer structure, such as steps or bullet points
• combine the best parts of two or more responses

Example of an iteration

This iteration adds user context, specifying that the content is for a busy parent who needs to complete the process quickly. Including this information helps the LLM adjust the tone, focus on essential steps, and make the guidance easier to follow for the intended audience. It also retains the previous improvements: clearer list structure, simplified language and shorter sentences to improve readability and accessibility.

Each time you adjust your prompt, note what worked better. After each interaction you should:

• review what the model generates
• identify what works and what does not
• give clear feedback and additional instructions
• repeat until the text meets your standards

Avoiding hallucinations

Sometimes LLMs 'hallucinate'. This means they make up answers that sound right but are not true.

To reduce this:

• give clear, specific instructions
• include relevant data or examples in your prompt
• tell the LLM to respond with 'I do not know' if it is unsure

Always check facts against trusted sources before using the output.

Reviewing and editing the output

You are responsible for making sure the content meets the right standards. Treat the LLM’s response as a first draft. Check and edit it carefully.

Check for:

• accuracy – does it reflect the correct policy, process, or service details?
• clarity – is it easy to read and understand?
• style – does it follow the Scottish Government Digital style guide?
• accessibility – is the language inclusive and readable?

When reviewing, make sure:

• headings are short and only capitalise the first word
• sentences are active and direct
• paragraphs are short
• lists are used for steps or grouped information
• dates and numbers use the correct format
• links describe their destination (not ‘click here’)
• acronyms are explained the first time you use them
• tone is clear, respectful and focused on the user

If the content does not meet these checks, edit it yourself or update your prompt and try again.

Using LLMs responsibly

Using AI tools in public service content requires the same care and accountability as any other form of publishing.

You should:

• keep a record of your prompts and outputs
• always review and edit before publishing
• be transparent if you’ve used an LLM to create or edit content
• save final versions in your team’s shared space so others can learn from them