Doc Example

The Terminalfour AI Connector lets AI applications such as Claude, ChatGPT, Microsoft Copilot Studio or Cursor work directly with your Terminalfour system. You type a request in plain language, and the AI finds content, makes changes or runs reports on your behalf.

This guide is for people who already use an AI assistant for everyday tasks but have never thought much about how they ask. You don’t need to know anything about MCP or APIs. It covers:

• How the connector, the AI and your Terminalfour system fit together
• What separates a strong prompt from a weak one, with side-by-side examples
• Ready-to-adapt prompts for common Terminalfour tasks
• How to work safely, especially when making changes
• What to do when things don’t go as expected

If you haven’t set up a connection yet, start with Getting started and come back here.

How it works

The AI does the thinking; the connector does the reaching into Terminalfour; your connection’s permissions decide what is allowed.

1. You type a request in your AI application.
2. The AI application (Claude, ChatGPT, etc.) works out which Terminalfour actions it needs.
3. The AI Connector passes those actions to Terminalfour, within your connection’s permissions.
4. Your Terminalfour system returns the results, and the AI answers you or takes the next step.

Three things follow from this, and they shape how you should write prompts:

• The AI only knows what it can see. It can’t read your mind, your intranet or last week’s meeting notes. If context matters, put it in the prompt.
• The AI can only do what your connection allows. Capabilities depend on the API key behind your connection. A read-only connection can explore and report but can never create, update, delete or publish.
• The AI can make mistakes. It may misunderstand a request or guess wrong about which section you meant. Good prompts reduce guessing; good habits catch the rest.

What you can ask it to do

Depending on your connection’s permissions, the AI can work with:

Area	Typical tasks
Content	Find, read, create, update and review content items
Sections	Explore site structure, create or reorganise sections
Media	Find and inspect media library items, check alt text
Content types and layouts	Review or build content types, write content layout code
Lists and metadata	Review and manage lists and keywords
Page layouts, Handlebars helpers and partials	Inspect and write template code
Users and groups	Review users, groups and memberships
Publishing	Publish channels, sections and branches
Documentation	Answer questions about how Terminalfour works

Not sure what your connection can do? Ask it:

What Terminalfour capabilities are available to you, and is this connection read-only?

What makes a good prompt

A good prompt answers five questions: what you want to achieve, where in Terminalfour, what to look at or change, how you want the result, and what the AI must not do.

You won’t always need all five. But when a result disappoints, one of them is almost always missing.

Ingredient	The question it answers	Example phrase
Goal	Why are you asking?	“We’re preparing for an accessibility audit…”
Scope	Where in Terminalfour?	“…in the Research section (ID 4521) and its subsections…”
Task	What exactly should the AI do?	“…find every image with missing or generic alt text…”
Output	What should the result look like?	“…and give me a table with the page, image name and a suggested alt text.”
Guardrails	What must it not do?	“Don’t change anything yet.”

1. Say why, not just what

The AI makes dozens of small judgement calls. Knowing your goal helps it make the right ones.

Weak:

Check the course pages.

Better:

We’re reviewing our undergraduate course pages before the new recruitment cycle. Check each one for out-of-date entry requirements or references to 2024 or 2025.

The first prompt could mean spelling, broken links, layout or anything else. The second tells the AI exactly what “check” means.

2. Be specific about where

A typical Terminalfour site has thousands of sections and content items, and names are often reused. “The news page” might match a dozen places. Point the AI at the right spot using names, paths or IDs.

Weak:

Update the contact details on the admissions page.

Better:

In the section Undergraduate > Admissions > Contact us (section ID 1287), update the phone number in the “Contact details” content item from 01 234 5678 to 01 234 9999.

IDs are the most reliable way to identify something. You can see section and content IDs in the Terminalfour interface. If you don’t know an ID, ask the AI to find the item first and confirm it with you.

3. Describe the task in concrete steps

Vague verbs like “improve”, “fix” or “tidy up” leave the AI guessing. Say what a good outcome looks like.

Weak:

Improve the page titles in the Research section.

Better:

Review the page titles in the Research section from an SEO perspective. Flag titles that are longer than 60 characters, duplicate another title, or don’t mention the research topic. Suggest a replacement for each one you flag.

4. Ask for the output you need

Tell the AI how to present results. This saves time and makes results easier to check.

• “Give me a table with columns for page name, section ID, current title and suggested title.”
• “Summarise in five bullet points, then list the details underneath.”
• “Only show me the pages with problems, not the ones that passed.”
• “Write it in plain English for a non-technical audience.”

5. Set guardrails

Guardrails keep the AI inside the lines you draw. They matter most when your connection can make changes.

• “Don’t change anything yet. Just show me what you would change.”
• “Only work inside the Engineering section.”
• “Keep the existing tone and don’t change any headings.”
• “Stop and ask me if you’re unsure which item I mean.”

Putting it all together

Here is a prompt that uses all five ingredients:

We’re preparing for an accessibility audit next month (goal). Look at all content in the Student Services section and its subsections (scope). Find images with missing alt text, or alt text that just repeats the filename, such as “IMG_2034.jpg” (task). Give me a table with the page name, section ID, image name and a suggested alt text based on the surrounding content (output). Don’t make any changes. I’ll review your suggestions first (guardrails).

It looks long, but it takes about a minute to write and saves several rounds of back-and-forth.

Give examples of what you want

When tone or format matters, show the AI an example. It is often easier than describing it.

Write meta descriptions for the five pages in the Postgraduate > Scholarships section. Keep each under 155 characters and follow the style of this one, which we like: “Find out about scholarships for international postgraduate students, including eligibility, deadlines and how to apply.”

Example prompts for common tasks

The prompts below are starting points. Swap in your own section names, IDs and goals.

Finding and understanding content

Exploring is the safest way to start and a good way to learn what the AI can see.

Show me the structure of the Faculty of Science section, two levels deep. For each section, tell me how many content items it has.

Find all content items that mention “Clearing 2025”. List the section path and content ID for each one.

Which content type is used for the news articles in the News section? List its fields and tell me which are required.

Auditing and reviewing content

Audits are where the connector shines: the AI can read hundreds of items and summarise what it finds.

Review all course landing pages under Undergraduate > Courses and assess whether each one is written at around a grade 8 reading level. Give me a table of the five hardest to read, with a short explanation and two specific suggestions for each.

Find all pages in the Undergraduate Admissions section that haven’t been updated in the last 12 months. Group them by subsection and flag any that mention dates or deadlines.

Check the Research section for content items with empty or missing summary fields. Just list them; don’t fill them in.

Creating content

Give the AI the content type, the location and the actual content, or the source it should draw from.

Weak:

Make a new event page.

Better:

Create a new content item in the section Events > Open Days (ID 2210) using the “Event” content type. Title: “Postgraduate Open Evening”. Date: 12 November 2026, 5pm–8pm. Location: Main Library, Room 1.02. Description: a short, welcoming paragraph inviting prospective postgraduate students to meet academic staff and tour the campus. Save it as a draft and don’t publish it.

Updating content

For edits, name the exact item and the exact change. Ask to see the result before it is saved when the change matters.

In content item 45872, replace every mention of “Dr Jane Smith, Head of School” with “Prof. Amir Khan, Head of School”. Show me the before and after text first and wait for my OK.

Bulk changes

Never start a bulk change with the whole site. Preview, test on one item, then expand.

1. Preview: “Find all content items in the Engineering section that link to the old URL www.example.ac.uk/apply-now. List them but don’t change anything.”
2. Test on one: “Update just the first one on that list to use www.example.ac.uk/apply instead. Show me what changed.”
3. Expand: “That looks right. Now do the same for the rest of the list and give me a summary of every item you changed.”

Site structure

I want to add a new subsection called “Sustainability” under About us (section ID 1002), with three child sections: “Our commitments”, “Reports” and “Get involved”. Before you create anything, show me the proposed structure and the section names and URLs you’ll use.

Media

List all images in the media library category “Campus photography” that were uploaded before 2022. For each, tell me whether it has alt text.

Content types and layouts

Work with me to create a new content type for our department landing pages. It needs a hero image, a heading, an introduction, up to six feature cards (each with an image, title, short text and link) and a contact block. Suggest the fields and element types first. Don’t create anything until I’ve agreed the design.

Show me the text/html content layout for the “Profile” content type and explain in plain English what each part does.

Users and groups

List the members of the “Marketing Editors” group and tell me which of them haven’t logged in during the last 6 months. Don’t change any accounts.

Handlebars and code

Help me create a Handlebars custom helper that formats a date as “Tuesday 12 November 2026”. Explain how it works and show me how to call it from a content layout.

Working safely

Look first, change second: ask the AI to show you what it will do, check it, then let it act.

The AI works quickly and confidently, including when it has misunderstood you. These habits keep you in control.

Read before you write

Split a change into two prompts: one to find and preview, one to apply. Adding “don’t change anything yet” to the first prompt is the simplest safety habit you can build.

Show me every content item you would change, and the exact change for each. Don’t make any changes yet.

Then, once you’ve checked the list:

Go ahead with items 1 to 4. Skip item 5; I’ll handle that one myself.

Confirm destructive actions properly

The connector always asks for explicit confirmation before destructive actions such as deleting content or removing fields from a content type. Your AI application will pass that question to you. Read it before you say yes. If you’re not sure what will be affected, ask:

Before I confirm, list exactly what will be deleted and whether anything else links to it.

Start small

A request like “update every page on the website” can have a much larger impact than you expect. Start with one section, or even one item, and expand once you’ve seen the result.

Pay extra attention to high-impact areas

Review changes especially carefully when they involve:

• Many content items at once
• Site structure
• Metadata and lists
• Permissions, users and groups
• Content types and layouts, which can affect every page that uses them
• Publishing

Use a read-only connection when you only need to look

A read-only connection can browse, inspect, report and answer questions, but any attempt to create, update, delete or publish is refused by the connector. This applies whatever the API key would otherwise allow. It’s ideal for audits, reports, demos and learning. Your administrator can switch a connection between read-only and read-write at any time.

For day-to-day work, ask your administrator for a connection whose API key covers only the parts of Terminalfour you need.

Check the facts

When the AI writes content, it may invent details that sound plausible, such as fees, dates, names or entry requirements. Always check facts against a trusted source before publishing. Tell the AI where the facts should come from:

Only use information that’s already in the existing course page. If something is missing, leave a placeholder like [CHECK: fee] rather than guessing.

Keep secrets out of prompts

Never paste passwords, API keys or other credentials into a prompt. The connection handles authentication for you. Be thoughtful about personal data too. Terminalfour doesn’t store or train on data sent through the connector, but your AI provider has its own policies, so check them.

Working in a conversation

Treat the AI like a capable colleague: brief it, check its work, and give it feedback rather than starting again from scratch.

Build up in steps

You don’t have to get everything into the first prompt. Complex jobs often go better as a short conversation:

1. “Show me the sections under Undergraduate > Courses.”
2. “Which of those use the ‘Course overview’ content type?”
3. “For those, list any that are missing a fees field value.”
4. “Draft wording for the three History courses, using the fees on the History department page.”

Each step lets you check the AI is on the right track before it goes further.

Correct it clearly

If the AI gets something wrong, say what was wrong and what you want instead. Vague feedback like “that’s not right” makes it guess again.

Weak:

No, try again.

Better:

That’s the wrong section. I meant the Admissions section under Postgraduate (ID 3310), not Undergraduate. Also, keep the descriptions under 155 characters.

Ask it to explain itself

If a result surprises you, ask how the AI got there. It will tell you which items it looked at and what it did.

Which sections did you search to produce that list? Did you include subsections?

What exactly did you change in content item 45872?

Let it ask you questions

For bigger tasks, invite the AI to check its understanding before it starts:

Before you start, ask me any questions you need answered to do this well.

Know when to start a new conversation

Long conversations can drift. The AI may start mixing up earlier instructions or sections. Start a fresh conversation when you switch to an unrelated task, or when the AI keeps repeating a mistake. Include a short summary of anything important from the old conversation.

Reuse prompts that work

When a prompt gives you a great result, save it. Many teams keep a shared library of tried-and-tested prompts for recurring jobs such as monthly content audits or new event pages.

Troubleshooting and common pitfalls

Most problems come from a missing ingredient in the prompt or a limit on the connection. This table covers the ones people hit most often.

What you see	Likely cause	What to try
The AI says it can’t do something	The capability isn’t available, or the connection is read-only	Ask “What Terminalfour capabilities are available to you, and is this connection read-only?” Then check the API key’s permissions with your administrator.
It worked on the wrong section or item	The name matched more than one place	Use the section path or ID. Ask it to confirm the item before acting.
It only looked at part of the site	Scope was unclear, or it didn’t include subsections	Say “including all subsections”, or ask “Which sections did you check?”
The answer is too generic	The goal and success criteria were missing	Explain why you’re asking and what a good result looks like.
It made up facts, dates or names	It filled gaps instead of asking	Tell it to use only existing content and to mark gaps with a placeholder.
The result is hard to scan	No output format was requested	Ask for a table, a list, or a summary followed by details.
It changed more than you wanted	No guardrails were set	Always preview first. Say “Don’t change anything yet.”
It keeps repeating the same mistake	The conversation has drifted	Start a new conversation with a clear, complete prompt.
The AI application can’t connect at all	Account, connection or configuration issue	Check your AI Connector account is active and the connection was set up as instructed. If it still fails, create a new connection or contact Terminalfour support.

A few habits to avoid

• One-word prompts. “Titles?” gives the AI almost nothing to work with.
• Asking for everything at once. “Audit the whole website and fix everything” is too big to check. Break it down.
• Accepting results without checking. Spot-check a few items in the Terminalfour interface, especially after changes.
• Assuming it remembers. A new conversation usually starts with no memory of earlier ones. Repeat the essential context.

Quick reference

Run through this checklist before you press Enter on anything important, or copy one of the templates below.

Prompt checklist

• Have I said why I’m asking?
• Have I said where: a section path, name or ID?
• Have I described what to do in concrete terms?
• Have I said how I want the result presented?
• Have I set guardrails, such as “don’t change anything yet”?
• For changes: will I preview before applying?
• For new content: have I told it where the facts should come from?

Templates

Replace the text in square brackets with your own details.

Explore

Show me [what you want to see] in [section name / ID], including subsections. Present it as [a table / a list / a tree].

Audit

We’re [goal]. Review [content type / pages] in [section name / ID] and its subsections. Flag any that [criteria]. Give me a table with [columns]. Don’t change anything.

Create

Create a new [content type] content item in [section name / ID]. Use these details: [field values]. Write [any text to generate] in a [tone] style, using only information from [source]. Save it but don’t publish it.

Update

In [content item / section ID], change [exact current text or value] to [new text or value]. Show me the before and after first and wait for my OK.

Bulk change

Find all [items] in [section name / ID] that [criteria]. List them with their IDs but don’t change anything. (Then: “Update just the first one.” Then: “Now do the rest and summarise what you changed.”)

Build

Work with me to design [a content type / layout / helper] for [purpose]. It needs [requirements]. Suggest an approach first and don’t create anything until I agree.

Further reading

• AI Connector documentation: setup, permissions, read-only connections and FAQs