This article sits alongside the slides from the Yorke Peninsula special session on AI for creative workflows. If you were in the room, this fills in the reasoning behind what got demonstrated. If you’ve been handed the slides without the two hours that went with them, it stands on its own.
This session ran differently to a standard Stage 1 delivery. It assumed you already had some AI foundation and pushed further into the specific territory of images, video, brand content and the tools that actually produce it. That’s the ground this article covers.
Additional event resources, discount links and extra slides available here
Not everything you’re doing needs AI
Before any tool gets opened, it’s worth sorting what actually needs AI from what’s already handled by something else. Financial software is the clearest example. Integrating AI directly with invoicing or accounting tools is fiddly to set up and, more importantly, not somewhere trust should be stretched thin. If you’re running Xero, use the AI that’s already built into Xero rather than bolting on something external. The same goes for anything that behaves like a CRM task: sharing an invoice with a client, filing it in a shared drive, following a fixed, repeatable process. If a task needs to be strict and repeatable with no deviation, that’s usually a systems problem, not an AI problem. AI earns its place when the task involves summarising, expanding, deducing or creating something that doesn’t follow a fixed template. Knowing which category a task falls into first saves a lot of wasted setup later.
Making AI sound like you, not like AI
The single most practical thing demonstrated in this session was building a humanizer skill: a text file, written in plain language, that tells your AI tool how you actually speak so its output stops sounding like a press release. It lives as a markdown file you can open and edit yourself, no coding required. Inside it you set out the small things: do you sign off with “hi” or “hello”, do you use “dear”, how do you actually phrase things when you’re not performing for an audience.
Once it exists, you don’t rebuild it. You test it, notice what’s still off, and give it a plain instruction to fix that one thing. It doesn’t need the whole file rewritten each time.
Even with a humanizer running, some habits are stubborn. Two personal instructions are worth setting up early, because they hold true almost everywhere: tell it explicitly which brand or client account it’s working on if you run more than one, and correct it if it gets confused about time zones or thinks something scheduled for today already happened yesterday. Neither is a big fix. Both stop small errors from becoming a pattern you have to keep catching by hand.
One honest caveat worth passing on: even with direct instructions not to use em dashes, and not to use American spelling, the tools still slip. It isn’t a sign the instruction failed. It’s closer to hiring a person: they arrive with habits of their own, and you keep correcting the same handful of things until the correction sticks. Building a standing “pet peeves” skill, a short list of the specific things you never want to see, and telling the tool to check against it every time, is the fix that actually holds.
Skills, connectors, and the risk worth knowing about
A skill only runs when you tell it to, but once it’s set up it sits available across everything, callable with a forward slash or just by naming it in a sentence. Different clients or different brands need different versions rather than one skill trying to cover everyone. Same underlying structure, different names, different specifics.
Connectors are what let Claude reach outside itself into the tools you already use. Some connect with a couple of clicks through a browse-and-add menu. Others, generally newer or more niche tools, need a custom connector: you paste in two lines the tool provides, under an MCP connection, and that’s the entire setup. It looks intimidating until you’ve done it once, and then it’s simply copy and paste in the right place.
The reason to slow down here, rather than install every skill going: prompt injection is a real risk, not a theoretical one. Early this year, a widely used AI assistant was found to have been manipulated through an email containing instructions invisible to a human reader, white text on a white background, that told it to gather information from across the connected system and send it to an outside address. It happened without a single click from anyone, and it was only caught because of an unusual spike in outbound data. Most hacks need some human error somewhere in the chain. This one didn’t. The practical response isn’t to avoid connectors. It’s to be selective about where skills and instructions come from, and where you’re not certain, run new material through a second AI tool first and ask it to assess the content for anything resembling a hidden instruction before it goes anywhere near your main setup.
Choosing an image tool: what actually matters
Several tools got tested side by side rather than picked on reputation. Adobe Firefly, included at no extra cost with an Acrobat Pro subscription, handles product image iterations reasonably well: upload an existing product shot, even a small, low-resolution file, and it can generate variations without a full reshoot. It isn’t flawless. The prompt-enhance feature can quietly insert details you never asked for, and it’s worth checking exactly what it added before accepting the result.
Gemini’s Nano Banana model does comparable work without extra payment if you’re already on Google Workspace, and tends to hold texture and small imperfections a little better, though bottle and object proportions can drift. Leonardo stands out less for raw capability and more for the boring but decisive factor: billing and cancellation are straightforward, it offers a genuine free daily allowance without needing payment details, and reviews on that front are consistently good. That matters more than it sounds. A tool with impressive output and a nightmare of a subscription to escape isn’t a tool worth building a workflow around. Higgsfield AI (its character tool goes by the name Reloom) does character consistency better than most of the alternatives, but was left out of the main recommendations specifically because of reported billing frustration, which says something about how that criterion gets weighted here.
The pattern worth taking from this, regardless of which tool wins: test the same prompt across two or three platforms before committing to one, and weight the decision by what other users report about getting their money back out, not just what the demo images look like.
What AI still can’t do, and why that’s useful to know
AI generates content, but it doesn’t know what matters. Ask it to design promotional material and it behaves like a keen junior: it tries to include everything, because it has no instinct for which single thing should carry the message and which should be dropped. Deciding what to leave out is still a human judgement call, and it’s one of the more useful things a design background actually protects you from getting wrong.
Two other limits are worth setting expectations around before you rely on a tool for anything client-facing. Typeface and layout consistency across headings and sections remains stubbornly difficult to hold, no matter how specifically you spell out size, colour and font for each element. And spatial or directional instructions, left of the building versus right, clockwise versus anti-clockwise, north-east versus south-west, are a known weak point. One attendee’s example from this session: a wayfinding diagram that needed an arrow pointing from the west side of a building to the entrance took roughly twenty attempts and still couldn’t be corrected reliably, eventually solved by stripping it back to a plain, hand-drawn-style line diagram instead of asking the tool to reason about direction at all.
None of this makes the tools less useful. It means the value is highest where the task is iteration and variation on something you’ve already directed, and lowest where the task needs a decision about hierarchy, direction or restraint.
The 90/10 rule
The most reliable working principle to come out of this session: let AI carry the bulk of the volume, and keep the last, most visible ten per cent under manual control. Applied to photography, that meant using AI for the surrounding elements or the iteration, while leaving the parts a viewer would recognise instantly as wrong, most often faces, deliberately close to the original or lightly blurred, because we know our own faces well enough that even a small distortion reads as unsettling. It’s a workable trade-off rather than a compromise: plan in advance for which specific part of an image is likely to be difficult, and build the shot so that part matters less to the outcome, or stays untouched.
The same principle extends to what content actually goes out under a brand’s name. Content that’s obviously AI-assisted and not pretending otherwise can perform perfectly well. Content trying to pass as fully authentic and not quite landing does real damage to trust, because the imperfect, unmistakably human version is what stands out and performs better once AI content becomes common and cheap. If in doubt, the honest phone photo taken at a bad angle beats the synthetic image that’s almost, but not quite, convincing.
Design is shifting, and it’s not really about AI
Worth naming separately from anything AI-related: the visual trend line is moving. The last decade or so pushed everything toward pastel, beige and muted tones. That’s now reading as dated to a younger audience, and bold colour, visible personality and a bit of imperfection are coming back, plainest in the return of a single striking image with bold yellow text over the top, no carousel, no elaborate layout. None of this requires complex design software. Canva handles it well precisely because the trend itself has simplified: less layout, more impact from a single strong choice.
The harder judgement call is timing. A brand that’s well built, with clear values and consistent communication, shouldn’t be knocked off course by chasing every micro-trend, and generally isn’t. The risk sits with adopting a trend so late that it reads as catching up rather than leading, the kind of lag a market notices even when the individual execution is fine. The right question isn’t whether a trend suits your taste. It’s whether it suits what your actual audience is looking for, which is very often a different question entirely.
Bringing scheduling, content and chat into one place
A late but exciting find for this session was Blotato: a single tool that connects multiple brands, multiple social accounts, image and content generation, and scheduling, without the friction of moving between five separate platforms to get one post live. Each brand can carry its own assets, tone and settings, so a business managing more than one identity, or an agency managing several clients, isn’t juggling logins to keep them distinct. It’s still early days of testing here, and it has rough edges (a draft can show as scheduled when it’s already gone out, for instance), but the scheduling, commenting, and multi-account management alone make it a serious contender to replace juggling native platform tools directly.
The practical workflow that held up best in testing: build the image separately in whatever tool gives the most control (Canva, for most day-to-day content), then bring the finished asset into Blotato purely for scheduling and distribution, rather than trying to get one tool to do everything end to end.
Where this leaves you
Test two or three image tools against the same prompt before choosing one, and weight the decision by billing experience and reviews, not just output quality. Build one humanizer skill and let it run in the background rather than editing every draft by hand. Be deliberate about what percentage of any piece of content is genuinely AI-produced versus manually held, and treat that split as a decision, not an accident. And before installing a skill or resource from outside your own tools, run it past a second AI system and ask it to check for anything resembling a hidden instruction.
None of this is about using AI less. It’s about knowing precisely which ten per cent of the work still needs a human hand on it, and making sure that ten per cent doesn’t get skipped just because the other ninety got fast.
