👋 Greetings {{first_name|earthlings}},

Emma here, co-founder of Spark AI. Everyone's talking about dogfooding at the moment – tech's least appetising piece of jargon (meaning using your own product internally)*. 

It came up all morning at an OpenAI Women in AI event I was at in London a couple of weeks ago. The operations lead at ElevenLabs described how a team of roughly 700, doubling every six months (with 500 roles open right now!), keeps everyone on the same page: an agent reads their Slack update channels twice a week, writes a script, cross-checks it against the source, turns it into audio in their own product's voices and posts it as an internal podcast people listen to on the commute. Candidates even talk to AI versions of their recruiters before the first screening call – and they've pointed the same technology at their own research, running more than 230 customer interviews in 24 hours with a voice agent.

So today we are sharing Spark’s version – with a twist, because we don't have an AI product to feed ourselves. We're the ones helping agencies and brand teams get commercial value out of AI, so our dogfooding is running our own process on ourselves. Last week our central team – the people who run Spark day to day, across operations, sales and marketing, rather than the AI coaches who deliver our AI programmes with clients – stepped out of the business for a day to do exactly that.

And rather than me telling you how it went, I’m handing the pen to Chris Caira, our Client Operations Director. Chris joined us ten weeks ago, one of a couple of brilliant new faces on our central team. New enough to see the practice with fresh eyes, close enough to the detail to tell you what actually sticks. And as always, there's a practical exercise to try with your team enclosed below. Over to Chris.

The team has grown quickly in recent months, and it's easy for growth to break the things you assumed were shared: the guidelines for how we use AI, the infrastructure behind it – shared AI projects, agents and skills (saved, reusable instructions the AI can pick up for a repeated task). A client CEO summed up the exam question on a call recently: the hard part is moving from AI for me to AI for we. Individually, everyone was making progress. Collectively, we were at risk of each running our own private version of it – separate prompts, separate habits, separate shortcuts nobody else could see.

Our internal AI practice at Spark runs on a fixed cadence. Protected Friday afternoons to create, learn and build. A fortnightly half-hour where we show each other what we've done and plan the next fortnight, working in SWAT teams on internal projects. And a full day each quarter to step back, reset and re-align as the technology moves. This was that day – the first with the new, bigger team in the room.

The aim was to raise the floor: everyone, not just the more experienced few, getting proper value from the tools rather than settling for the 10% you get from treating AI like a chatbot. It's the exact thing we tell clients. This time it was aimed at us.

The shape of the day matters as much as the content, so here it is, step by step.

1. Whole team, out of client work, in deliberately mixed pairs. A full day, everyone, no exceptions – not a lunch-and-learn squeezed around delivery. And the pairs matter: mix backgrounds, perspectives and skills. Put the operations brain next to the creative one, the longest-serving next to the newest joiner. Two people with different instincts looking at the same problem build better things than two who think alike – and the knowledge spreads in both directions.

2. Start with quick wins everyone takes home. First hour, get the quick wins out there: things that pay for themselves the same week. A morning brief that runs on a schedule before you're at your desk. Personal instructions so the AI stops flattering you and starts challenging you. Dictation (Wispr Flow 😍) , because you'll give ten times more context speaking than typing. Setting up your AI Chief Of Staff to brief you on the day and kick off tasks. These fundamentals are always where the room switches on, and even once you have them set up they need continuous revisiting and refreshing. Alice had been out of office for two days – her first morning brief came back with what she'd have spent half the morning piecing together from her inbox, including that Jules had covered a client negotiation on her behalf and sent over the training breakdown. "That would have saved me half an hour," she said, looking slightly annoyed she hadn't been running it all along.

3. Reinforce the rule before the tools. Active minds, not autopilot. Form your own view first. Brief the AI well. Then interrogate what comes back. If you don't know what's in the output, you've outsourced too much of your own thinking. AI can generate a 10-page document for something that only needs 1. The AI can write it quickly, but nobody can read it quickly. If you skip the 'is this good?' step, the time you've saved has just been lost by whoever has to wade through the output. Make sure AI is the middle step, never the last one. Every exercise for the rest of the day ran through that filter.

4. Audit your shared foundations. One source of truth for company context – services, pricing, who your clients are, how you sound – with a named owner for this shared context and a review rhythm. This is the first thing to sprawl as a team grows, so we go looking for it. Sure enough, we found three slightly different versions of the same context in three places, exactly the drift we warn clients about. Caught, and cut back to one and the new owner of it appointed. Unglamorous, and probably the highest-value fix of the day.

5. Explore before you build. There's a library of pre-built skills sitting in these tools that most teams never open. And they are excellent! So before anyone built anything, everyone picked one and tested it on real work. Alice ran an account research skill on a real prospect ahead of a call this week and found it more methodical than she'd ever be by hand. Asta used the document co-author to turn what we know into a detailed customer journey map. Then, and only then, build your own – and the trick is to do the task once, conversationally, until the output is right, then save the whole process as a reusable skill. That's how Jules built one that read his last hundred sent emails, worked out on its own that he writes differently to clients than to the team, and produced a style guide for each.

6. Leave with names against things. This is the step most teams skip, and it's the one that decides whether the day was a reset or a jolly. We left with a shared backlog of what we're building next, a name against each item, a single owner for the skills library and its version control, and the next quarter's cadence back in everyone's diaries – including the people who weren't here for the last one.

We're not starting from one place. With a bigger team, we're not all at the same level, and that's what happens when you bring new people in. The core team works with AI fluently, as you'd expect from us; our newest joiners are earlier in the climb. Knowing exactly where each person sits is what lets us pair deliberately next quarter, the most fluent alongside the newest, so people come up fast rather than falling behind without us noticing.

Not everything landed straight away either. Onboarding new people has its friction. A couple hadn't finished setting up their tools before the day, so their first 30 minutes was admin (running /setup_cowork) than learning. If you're bringing new joiners into a practice like this, send the setup checklist out the week before. 

And the ceiling is higher than it was a year ago. In one day we saw interactive proposal pages built in a chat window, a customer journey map assembled from knowledge that previously lived in six people's heads, a live marketing analytics dashboard, a buyer personas drawn from a hundred discovery-call transcripts, and an inbox swept for actions missed over the past five days. All of it built by a team with no technical background.

Three updates worth knowing – and the second one makes step five of the walkthrough even easier.

Google has just launched Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, models built to use fewer tokens per task while still holding up for everyday work. This drops right in the middle of a wider conversation about token costs, as models keep improving and getting heavier to run, and it's putting real pressure on the leading labs to respond. Anthropic has settled this for its own priciest model by including Fable 5 permanently in Max and Team Premium plans, capped at half your usual usage limit. OpenAI has taken a different approach, temporarily lifting the five-hour usage cap on ChatGPT Work and Codex, with quotas being reset regularly since. It's a good reminder to always check what model you're using against what the task needs – you don't always need the most expensive one.

Read more: Google's announcement

Anthropic shipped four things in quick succession. "Record a Skill" landed in Claude Cowork (the version of Claude that works in your files, not just the chat window) on 21 July: do a task once on screen while narrating what you're doing, and Claude turns the recording into a reusable skill it can run again on its own. You'll find it under the + menu in the desktop app – it's not on web or mobile yet, and it's rolling out gradually, so update your app if you can't see it. Read more: How to create custom skills

Also, Claude Opus 5 is now generally available, with better code, agents that run reliably for hours and reasoning that holds up across long documents, matching the top-end Fable 5 in many areas at half the price – plus an "effort" dial to turn thinking power up for hard tasks and down for routine ones. 

Next, from 3 August, Cowork runs on web and mobile too, with sessions running in the cloud: start a task at your desk, check it from your phone, and scheduled tasks keep running with the laptop shut. 

Lastly, Claude can now act in all your Microsoft 365 apps – sending emails, managing calendars, updating SharePoint files. This one doesn't work out of the box: your admin has to enable Claude and switch on each write permission individually, choosing "Allow" or "Ask" per action. We like to set it to "Ask", and you can also only give it access to a specific folder for outputs for an additional safety net. 

That was a lot, phew!

Google renamed its brilliant research tool NotebookLM to Gemini Notebook on 16 July, bringing it fully under the Gemini brand with a matching logo. It's the same product underneath – your existing notebooks and shared links carry over automatically, so nothing needs rebuilding. An exciting addition is that every notebook can now write and run its own code against the sources you've uploaded, so instead of just reading and summarising your documents, it can analyse and generate outputs directly from them. Live now for the top subscription tier, rolling out more widely over the coming weeks.

Read more: Google's announcement

Well that’s been quite the newsletter, so its over and out from us today! Until next time,

Chris and Emma 

*Apparently (according to Claude) the term “dogfooding” took hold at Microsoft in 1988, when a manager sent an email titled "Eating our own Dogfood", challenging the team to actually run the software they were building. He'd borrowed the phrase from a 1970s Alpo advert in which Lorne Greene fed Alpo to his own dogs – and the president of a rival pet food company reportedly went one better, eating a tin of it at shareholder meetings. Yikes.

Spark AI is the AI performance partner that builds organisation-side AI capability for agencies and brand teams. We do this through AI training, consultancy and agent building.

We’ve worked with 80+ agencies, published the #1 bestselling book on AI for agencies and publish the AI in agencies benchmark, and teach at Oxford University.