One paragraph, 55 minutes: why interactive guides are due a comeback in the age of AI

The Brief

• The result: A brief of about 80 words, pointed at our six-part AI Agent Series, produced an illustrated, interactive companion in 55 minutes. Checking its 504 claims took the rest of the day: none false, 11 with a hedge or a detail lost.
• What we did: We borrowed a great brief from Anthropic’s Boris Cherny and applied it, with minor changes, to the six published parts of our series in the Claude app, checked the result with a second model and our claim ledger, then gave Claude one more prompt for a throwaway free-kick game.
• The takeaway: Making it took one paragraph. Deciding what it should say, checking it and having the right to publish it took the rest of the day. We gained time and quick, impressive results, and lost the spirit of collaboration I relished in my early days working on interactive formats.

The Story

When interactives took months

I’ve been fascinated by interactives since 2014, when I worked at the BBC on the newly launched iWonder guides. Each one started with a question people were curious about (described internally as “the curiosity principle”) and worked through it in short steps of text, video, audio or a game.

Match of the Day had secured the FA Cup rights, so we made the most of it with maths guides built around the competition. The first was on the curve of a free kick and the parabolas behind it, with videos featuring Dan Walker, Rachel Riley and Robbie Savage (I also wrote the scripts, and it was amazing to hear these talented people read my words aloud!).

This kick-off guide (pun fully intended) was showcased on the BBC Sport homepage and reached hundreds of thousands of people, far more than we expected, which felt like a sign there was something to the format. I produced and wrote it with the iWonder team and Aer Studios, who built “scrubbables” you could drag through frame by frame (an idea I later took to Red Bull, with a scrubber of the astounding champion cliff diver Gary Hunt; the link goes to my article, though the interactive itself was lost in a later site redesign).

For the guides, I worked with a wonderful and diverse team of experienced pros, including senior sports journalists, technical wizards, SEO whiz kids and respected educational experts. Together we shaped each one into something worthy of the BBC’s main pages, and I learned a huge amount in a relatively short time.

Each guide took months and a full team to finish. iWonder launched as a flagship, and two years later the BBC closed it, redeploying its formats across BBC Online as part of a plan to save £15m.

There was clearly an appetite for this kind of online format at the time. In 2014 I also interviewed at BuzzFeed for a job as an interactive journalist, and pitched a map for the coming UK general election whose features shifted towards a stereotypical left- or right-leaning voter, depending on where you looked. It would have needed months with a developer and a design team too. I didn’t get the job, which may tell you something about the map.

I thought about both this month, watching Claude build something similar from one paragraph in under an hour.

What Cherny did

Earlier this month Boris Cherny, who created Claude Code at Anthropic, shared an “interactive website companion” to Acquired’s episode on The Home Depot, built by Claude, along with the result and the prompt. That’s Anthropic’s model, demonstrated by an Anthropic leader, on a show where Claude is a season partner. So we ran the same brief on something of our own, and checked everything that came back.

His prompt reads like a good creative brief. It says what the thing is for (the “business wisdom” in the episode), gives two references to aim at (“New Yorker or nytimes data viz vibes but better”), grants permission to keep going (“use lots of tokens and iterate till you’re proud of it”) and adds one technical hint. Comms people write briefs like this for designers all the time.

For our Applied experiment, I changed three things: the source (our six-part series), the purpose (practical wisdom for communications teams), and I dropped “historical photos”, because other people’s photographs are a rights question we’d rather not create.

What came back

I pasted the brief at 07:12 on Tuesday 14 October. At 08:07 there was a published page, “The Human Share” (not my title!), with 21 watercolour illustrations painted in code and a bar that tracks the agent’s share of the work against the people’s as you scroll. Before publishing, it sent a second agent to check its own draft against the six articles, and fixed 15 loose paraphrases and one false label.

Skip past the companion

Scroll it here, or open it full screen.

Then came the checking, which is where the day went. ChatGPT reviewed the page three times, and its second review found real bugs that Claude’s own tests had missed, because those tests only covered the cases Claude had designed for. Our claim ledger then went through 504 claims: none false, 11 where a hedge had dropped out or a detail had attached to the wrong thing (a warning period of “six weeks before a change” where the series says six weeks before the change was announced). All 133 quotations were word for word. It brought in no outside data, and the things it invented, from the scrolling bar to a “where to look first” ranking the series never makes, are labelled as its own.

Two honest limits. It had a more solid standing start than Cherny’s run may have had, thanks to using our own previously developed and published content, inside our Comms With AI project with its instructions and memory to hand. And I didn’t measure what it cost.

And one more thing, because why not…

Then, as a throwaway, I gave Claude one prompt for a game, a nod to that FA Cup guide: bend a free kick round a wall, with the curve drawn after every shot. Twelve minutes later (I was changing my daughter’s nappy for most of them…) it looked finished and worked for exactly one kick, then froze. Aiming only went hard left or right, because it expected a pull-back slingshot and on a phone you drag towards the goal. Claude hadn’t spotted either problem.

Not to worry! I played it once on the laptop and once on my phone, sent two screenshots and a few voice notes, and three minutes later it was a game, with a short explainer underneath on why the ball bends. It’s hard: I’ve scored none in ten. Claude then ran the game’s own physics on nearly 47,000 shots and found that about one in six go in, mostly aimed well wide of the left post and curled back, which sadly suggests that, my natural position being goalkeeper, the problem must be my free kicks…

Have a go.

Skip past the game

Play it here, or open it full screen.

Why your website, and good agencies, matter again

Fewer than one in three US Google searches now ends with a click on a link, according to research reported in the UCLA Anderson Review. That’s a big part of why people say websites are dying.

I think this run points the other way. A post on a social network can only do what the platform allows. A page you own can now do almost anything you can describe, and the audience, the data and the rules stay with you.

The same goes for the tools you build for clients. Carly Martinetti, who works on PR and AI visibility for mid-market companies, posted last week about a client who handed her a wish list of publications. She spent three days building a dashboard that maps each one to opportunities for the client, surfaces the relevant detail from the client’s knowledge base and shows who to pitch, with the research re-run every week “so the list never gets stale”.

The big platforms are building this in, too. Claude Dashboards, in beta, turns plain-language questions about your company’s data into dashboards; OpenAI has a guide to building one in ChatGPT that stays up to date; and Google’s Sheets canvas builds them from a spreadsheet. Ethan Mollick launched his new book, Co-Existence, last week on a site that, in his words, “Opus 5.5 redesigned. I told it to make it better.” I suspect a static article will start to feel stale next to pages like these, though there will always be a place for straight text in longer essays and books.

Being able to build something like this in an hour shows anyone what’s possible, and it makes for a much better brief. Taking it further is where experienced agencies and studios come in: they know how to make an experience work for a real audience, at scale, and keep it working. That’s where I’d take a prototype like ours next.

The catch is the one the run made obvious. Once making is cheap, the work moves to the boundaries: a strict CMS that controls what can go live, checks against the source, and clarity about whose material it is. That’s comms work, and it’s where I’d spend the time saved. (Our Claims Substantiation Checklist is a good place to start on the checking.)

The Practice: what you can put into practice today

1. Pick one long thing you own outright (an annual review, a report series, a recording nobody finishes) and run the brief on it this week. Ours, ready to adapt: “The six-part AI Agent Series on Applied Comms AI (appliedcomms.ai) is really useful. I want you to make an artifact that narrates the entire series. Make it beautiful and interactive and focus on communicating the nuggets of practical wisdom for communications teams throughout the series. Make it really good, use lots of tokens and iterate till you’re proud of it. Maybe add watercolour illustrations throughout, aim for New Yorker or nytimes data viz vibes but better. I think you can use OpenCV for watercolour images”
2. Check it as you’d check copy: every figure and every attributed lesson against the source, and ideally a second model reading it cold.
3. Write the rule down: your own material, yes; someone else’s only with their permission; and always say how it was made. Then note what worked and what didn’t, and build it into the brief, or a more detailed prompt, next time.

Which long thing of yours, the one nobody finishes, would you give a companion?

Every Applied piece follows the same shape (The Brief, The Story, The Practice) and is made the same way: my argument and examples, drafted with Claude, every claim checked against its source, and a final edit by me.

The next Comms With AI Leader Interview is Zoe Amar on what charities do next with AI, on Wednesday 18 November at 16:00.