Content design · UX writing · Content strategy

I make complicated products make sense.

I'm Caroline, a content designer in Leeds. I write interface copy, onboarding and help content for regulated software, working with product, design, engineering and legal from early ideas through to launch.

45%fewer support calls in a year at IRIS, through content strategy and self-service design
26%fewer support tickets in six months at CUBE, from 500+ help articles built to one standard
51.19%response rate on the customer survey that shaped a new payments app
~200help articles a month, shipped by a team of three

Selected work

Regulatory intelligence product · New-user onboarding

New-user onboarding: copy, research and a clickable prototype in one package.

The problem

New users had to find their own way to value. The people shaping onboarding (Product, UX, Engineering and Customer Success) needed something concrete to react to, quickly.

The approach

I mapped the whole onboarding journey and delivered it as one package, without needing a Figma file.

  • The journey. Six in-app pieces built by Product and Engineering, and seven Pendo guides in three tracks: a welcome lightbox and product tour, four nudges towards a first feed, saved search and notification, and a reusable What's new tooltip for each release.
  • One source for every string. A copy deck covering every screen, the rules for each track, and a proposed measurement plan.
  • A clickable prototype. Every piece in context, with a Pendo-only view and a copy panel for demos.
  • Hand-offs ready to use. A design brief for the wizard, four stories' acceptance criteria reviewed and rewritten, and a Pendo build sheet so the guides can be built directly.

The decisions

  • Voice. I corrected the copy against the real screens and wrote it in the product's voice.
  • Sentence case
  • No "please"
  • Exact UI labels
  • Legal. A disclaimer went in after feedback from Customer Success. It is awaiting approval.
  • The compromise. The tour became a single 13-stop tour at a stakeholder's request. That is far longer than best practice, so the plan is to track where users leave and shorten it if most do.
  • Measurement. A holdout test, which needs several months at roughly 35 new visitors a week.

The result

The prototype let me share my ideas directly with the UX designer, who built them into his own. That version is now in build and ships in December.

In build, shipping in December

The words, in context

Guide 5 · Save a search tip · Pendo tooltip

Save this search as a feed

Turn this search into a feed. [Product] adds new matching content to Feeds, so you don't need to run the search again.

Not nowSave search as feed

The string as written. Product name removed.

What the string is doing

  • Right person, right moment. It only appears for people who have searched several times and never saved a search, on the results page, next to the button it describes.
  • Exact UI labels. It names the destination the way the navigation does: Feeds.
  • The benefit, in the user's terms. You don't need to run the search again.
  • A way out that isn't a refusal. Not now.
  • One goal, one event to tag. Search saved as a feed.
The clickable prototype: every guide in context, with its audience, trigger and goal alongside.
The copy deck: one source for every string, with a JSON export for the build.
  • 7 Pendo guides
  • 3 tracks
  • 1 copy deck for every string
  • UX writing
  • Prototyping

Concept prototype · Industrial equipment documentation

From a library of PDFs to a product experience.

On site: tasks for one exact unit (left), and a task with its warning before the first step (right).

The problem

A global manufacturer of cooling equipment publishes its manuals as print PDFs. The same manual is read by an installer on a roof, an operator in a plant room and a buyer at a desk, in different regions and languages. The brief asked for manuals ready for a help centre.

The approach

I built a working prototype of a documentation hub, in four views.

  • On site. Type the serial number or scan a QR code, and the phone shows only the tasks for that exact unit, saved for offline use.
  • Choosing equipment. Filter products by industry and compare two side by side, including which documentation each one has.
  • Content map. Content organised by the moments in an owner's life, from choosing to retrofit, with what exists, what is partly there and what is missing.
  • One source. A single topic producing both the PDF manual and the online help, adapted by product, region and language.

The decisions

  • Tasks, not documents. "Replace a fan belt" beats "Operation and maintenance manual, section 6". Each task is a short page with its own warnings, tools and parts.
  • Warnings come first. Safety content is reusable and owned, and sits before the step it applies to.
  • Region and language are separate. Region sets the units, the electrical supply and the standards. Language sets the words.
  • Honest about drafts. Translations are labelled as machine drafts that need a professional linguist.

The insight

Manuals are the start, not the whole job. Most of the gaps were about structure, publishing, data and ownership, not writing. One shared warning becomes 68 outputs across two products, two formats and 17 regional language versions. Fix it once and every one of them is fixed.

Try the prototype

Concept, not shipped

The other views

Select any image to enlarge it
Content map and gaps
One source, every output
Choosing equipment

The company is anonymised. Products, serial numbers and figures are invented examples.

  • Information architecture
  • Task-based content
  • Localisation
  • Mobile and offline
  • Prototyping

Client project · UK local government · Electoral management software

Help for the people who run elections, organised by the job, not the software.

Before: organised by software module

Module 5 Processing Absent Votes Processing AV synchronised from the portal

After: organised by the task

Absent votes Process online applications

The problem

A UK software company makes the system councils use to register voters, run the annual canvass and deliver elections. The help was a set of basic training guides, organised by software module. But nobody asks "How do I use Module 4?". They ask how to add a new elector, or where someone's polling station is, usually against a statutory deadline.

I started with access to the live system and those guides.

What I built

  • Discovery assets to understand the products, the users and the content that existed.
  • A content foundation. Five content principles, voice and tone for each context, and a style guide. Anything citizens read follows the GOV.UK style guide, aiming for a reading age of 9.
  • The help centre's information architecture. 16 sections and more than 300 topics, guides for seven roles, and a metadata model, so one topic can appear in several places without being duplicated.
  • Task-based help topics, rewritten from the training guides to GOV.UK content standards and Minto's pyramid principle: the answer first, then the steps. The first set covers the 11 enquiries customer services staff handle most.

The structure, on four principles

PrincipleFor usersFor the business
Task-ledGet the job done fasterFewer support contacts
Role-awareSee what's relevantBetter adoption
Search-firstFaster answersMore self-service
Module-supported underneathConsistent informationEasier maintenance

The decisions

  • Plain English is not dumbed-down English. Statutory terms like Individual Electoral Registration stay exactly as they are. The skill is explaining them on first use.
  • Role-aware entry points. A customer services user asks "Is this elector registered?". An electoral services manager needs to publish the register. Each gets their own way in.
  • Help that prevents mistakes. Topics tell staff what not to say, as well as what to select.
  • Built for peak periods. Help has to work hardest during the canvass, on polling day and on count night, because that's when it usually fails.

The words

A help topic for customer services staff

Check application status

Find out the current status of an elector's registration application.

Before you start

You must have permission to view elector and application information.

Check an application

  1. Search for the elector.
  2. Open the elector record.
  3. Review the application information.
Received
Waiting to be processed.
Further information required
Can't be completed until more information is provided.
Approved
Will be added to the register in line with the publication timetable.

What to tell the elector

Don't confirm registration until the application has been approved and published.

Shortened from the full topic. Product name removed.

Before and after, from the style guide

Status update to a citizen

Before: Your appeal is being processed. You will be contacted in due course.

After: We have received your appeal. You will hear from us by [date]. If you have questions in the meantime, contact [name] at [council name].

Error message

Before: Something went wrong. Please try again.

After: The file could not be uploaded. Check the file is in .CSV format and under 10MB, then try again.

Every citizen notification has to answer four questions: what has happened, what happens next, by when, and who to contact.

See everything I delivered

  • 16 sections
  • 300+ topics
  • 7 role guides
  • GOV.UK style guide
  • Information architecture
  • Style guide

RegTech · AI features · EU AI Act

AI features people can trust: what's AI, what's source, and where you decide.

One disclosure model, four tiers

  1. 1First useOnboarding, the first time someone opens the feature.
  2. 2Page footerA page-level notice, so cards don't repeat the full disclaimer.
  3. 3Output labelOn each AI result.AI ReasoningAI-assessed
  4. 4Decision pointNext to the confirm and review actions, where the person decides.

The problem

The product's customers are regulated financial institutions, and it was adding four AI features, called Coworkers. Each needed copy that tells people what the AI did, keeps the source visibly separate from what the AI produced, and puts the human decision where it belongs. The EU AI Act makes all of that a legal requirement as well as good design.

What I did

  • A tiered disclosure model. One universal template with slots for the source, the output type and the decision action, written against Articles 13, 14, 26 and 50 of the EU AI Act. It's now used across all four Coworkers.
  • Copy for every surface. Onboarding, page footers, output labels and decision-point copy for each Coworker, plus limitation language placed next to the outputs it applies to.
  • Screen reviews. One review caught a banned term, AI interactions that weren't disclosed, and metrics the product couldn't support. Issues first, then specific rewrites, with open decisions flagged to the right owner.
  • Accessibility built in. Text paired with every AI icon, screen-reader parity, and examples that work in any market.
Keeping source and AI apart: the Obligations panel

AI snapshot

Extracted content

AI-suggested classification BETA

Wireframe of the structure. Real content removed.

The outcome

  • Priorities is being rebuilt around user feedback. Strong findings from user testing drove a full overhaul and a change of direction. I'm rewriting the copy alongside it, and testing is going well.
  • One model, four features. Users see the right notice at the right moment, and every new Coworker starts from the same template.

Language decisions, settled on principle, not preference

DecisionInstead ofWhy
Priorities CoworkerScreening Agent"Agent" is banned in user-facing copy, and "Screening" names the mechanism. The Coworker ranks publications, which is more than screening them out. It also kept "screen" free for another Coworker being named at the same time.
Enforcements CoworkerEnforcement Action Coworker"Action" implies the user has to act.
ReviewedArchiveChosen through an accessibility and localisation analysis.
AI snapshotA compliance-first or marketing-first headingSettled after a debate between compliance clarity and marketability. It's clear about what produced it and still easy to find.
BETA, used sparinglyBETA on anything newOnly for a specific reason, such as a taxonomy that hasn't been validated. Sometimes "low confidence" is the more honest label.
12pt minimum textA 10pt exception for legal textLegal text has to be readable too. The accessibility floor held.

How I name AI features

  • Name the outcome, not the mechanism. Outcome-led names (Insights, Priorities, Results) over capability-led ones (Interpret, Evaluate, Prioritise).
  • No internal concepts. "Agent" and other anthropomorphic terms stay out of the interface.
  • Fit the family. Priorities, Enforcements and Obligations share one short, noun-led pattern.
  • Get it right early. First names tend to stick, so coherence comes before launch.

Disclosure model used across four AI features

  • 4 disclosure tiers
  • 4 AI features
  • EU AI Act Articles 13, 14, 26, 50
  • Naming
  • Accessibility

RegTech · EU AI Act · Working with Legal

From policy document to an organisation that checks itself.

The problem

Our customers are regulated financial institutions, and they will ask how we govern our own AI. AI-assisted prototyping was also getting features to engineering without any governance check. A team of three can't brief every squad one by one, and a document alone wouldn't change behaviour.

The approach

I built a governance kit, then built the organisation around it.

  • The kit. Five plain questions, asked in order, that turn the Act's obligations into practical guidance for product, design and engineering.
  • An executive sponsor. I made the case for why a RegTech company has to govern its own AI credibly, and secured senior sponsorship.
  • A champions programme. Volunteers across the business who run the review at the point where decisions are made.
  • Legal alignment. A briefing for counsel on how our agentic AI features should be classified.

The result

  • 8 champions across 3 dev squads, Legal, Customer Success, UX and Product.
  • Every agentic AI feature carries EU AI Act-appropriate legal wording.
  • A review gate. Nothing ships until the champions have reviewed and approved it.

Draft disclosure wording from the kit

  • AI-generated
  • Generated by AI. Review before use.
  • You are chatting with an AI assistant.

Sponsor secured, champions programme launched

Inside the kit

Select any image to enlarge it
Live in Confluence
The five questions
Who does what
One kit, every team
Design checklist
Assessment note
In Claude Code (illustrative run)
  • 8 champions
  • 7 teams covered
  • 1 review gate
  • Legal and compliance

RegTech · Content operations

Three people. Five products. About 200 articles a month.

The problem

Every product releases twice a month, and every release needs about 20 new or amended articles. Documentation, in-app guidance and release comms all have to land together, in more than one channel. Standards drift quickly at that pace, and content that suits a human reader can fail when an AI agent retrieves it.

Working harder wasn't an option, so I built systems.

The approach

One Claude project per product area, each pre-loaded with that product's standards, past decisions and voice rules. Anyone on the team can pick up any product mid-cycle and stay on voice. Content is authored once in MadCap Flare and published to both channels.

The result

The three of us (two PX designers and me) ship roughly 200 new or amended articles a month across help centres and Salesforce Knowledge, and tag every new feature in Pendo. No extra headcount.

  • Team of 3
  • 5 product areas
  • ~200 articles a month
  • 2 channels

51.19%

Survey response rate. The answers fed the product decisions for a new payments app.

IRIS Software · IRIS Pay

The survey that shaped a payments app.

Before building anything, we needed to understand how customers handled payments. A digital engagement survey drew a 51.19% response rate.

The result: the findings helped us develop and launch IRIS Pay, followed by the onboarding and UI copy that gets new users set up.

  • Research
  • Onboarding
  • UI copy
  • Payments

RegTech · Tone of voice

Hours of manual checks, now minutes.

At 200 articles a month, nobody can check each one by hand against a new tone of voice guide, least of all one with antitrust-sensitive language where mistakes carry real risk.

I built a reusable Claude skill that checks any content against the guide. I'm now adapting the guidelines for UX and the design system, so the interface meets the same standard as the help centre.

The result: every piece is held to the same standard instead of relying on a reviewer's memory. The time saved goes back into writing.

Skill live, design-system version in progress

Before

Before: Unmanaged, ungoverned and inconsistent content in multiple locations.


After

After: On brand, accessible and relevant in one place, for users and colleagues.

IRIS Software · Help Hub

30+ product help centres, one front door.

As Head of Content Design I brought scattered help content into two hubs, one for UK customers and one for global, hosting more than 30 product help centres.

The result: across IRIS, content strategy, clearer information architecture and self-service design cut support calls by 45% in the first year.

  • Information architecture
  • Governance
  • Self-service

RegTech · Support content

Release notes in, article drafts out, before Support has to ask.

Agentforce can only answer from articles that exist, and every release changes what customers ask. Support had to work out which articles each release needed and write them by hand.

Now a scheduled Claude task reads each week's release notes, decides which changes need a new article, drafts it in plain US English and posts it to Support's Teams channel. Every draft is marked pending review. A person publishes it.

The result: 88 article drafts written for Support across two products.

Three of the drafts, titled the way customers ask

  • Can I set zero advance days for an exempt notice?
  • Why are the notice type and effective date already filled in on my reinstatement or rescission notice?
  • Why can't I reset a password, reset MFA, or unlock a user who signs in with single sign-on?

Running weekly across two products

RegTech · Help content

A deflection strategy only works if the content underneath it does.

AI agents can't render everything a human reader can, such as markdown tables. I set Agentforce-optimised content standards, developed a ticket deflection strategy, and evaluated Intercom Fin, Agentforce and WalkMe against our publishing pipeline.

The result: the standards now shape the help content, with one source serving both readers. Customer Support own Agentforce, and I align the content with them.

Strategy prepared for executive sign-off

RegTech · Working with PMs

Content standards that travel with the prototype.

PMs began using Claude Code to produce UI prototypes, and the copy bypassed content review. With a team of three I couldn't review every prototype, and I didn't want to become the bottleneck.

So the standards go where the work is: a repo-level brief, a content checkpoint prompt and a Jira ticket creator. Enabler, not gatekeeper.

In development

RegTech · In-product help

Help content that answers questions inside the product.

However fast we publish, users still leave the product to find help. I designed an assistant that connects our MadCap Flare help output to the Claude API, so the content the team already writes becomes contextual answers in the product.

The result: I took it from prototype to a full engineering spec.

Prototype and spec complete

How I work

The process I set for my team at IRIS, and still use

Clear direction and vision

Purpose and objectives first. Everybody knows what we are here to do.

A design process that works

Define and communicate how we want to work. No more confusion.

Clear roles and deliverables

Know what needs to happen and when. Understand what comes next.

Research and iterate

Use data and feedback to learn, design and build. And don't stop there.

About me

I started writing for software in 1996 and I've been designing content since 2013. Teams, products and complex problems, solved with clear, purposeful language.

I've built and led a 14-person content design team across 37 SaaS products. These days I'm Product Experience Manager, leading a team of three and staying hands-on. I own in-product copy, Pendo guidance, help centres and release communications for a product suite where every product releases twice a month.

After years of building teams and systems, I want to spend my time on the part I love most: interesting work, solving real problems for real users.

I think end to end. I care about the full journey, not just the screen in front of me. And I believe the best content design is the kind users never notice, because everything just works.

I also volunteered as Documentation Manager for Team OSV's Fluid Heart app, which helped kidney failure patients track their weight during Covid-19 isolation.

My team right now, in numbers.
  • CUBE GlobalProduct Experience Manager. RegTech for regulated financial institutions.
  • iDealsSenior Content Designer. Web and mobile experiences for a new SaaS product.
  • IRIS Software GroupHead of Content Design. A 14-person team across 37 products.
  • dunnhumbyDesign Operations Manager. Style guide, voice and tone, design system rollout.
  • AnaplanTechnical Documentation Manager. A fortnightly release cadence through IPO-era growth.
  • EarlierTechnical documentation and communications roles.

View my CV (PDF)

Tools: Pendo, MadCap Flare, Salesforce Agentforce, Intercom Fin, WalkMe, Gainsight PX, Figma, Claude and Claude Code.

Let's talk.

If your product is complicated and your words need to be clear, I'd like to hear about it.