I've been saying for a while that the real skill with AI tools isn't picking a favourite — it's knowing which one to hand the pen to. This week gave me a clean, documented example of exactly that, so I thought I'd write it up. I wanted a Gemini "Gem" — a custom, reusable instruction set — that could generate the headline images for this blog on demand, in my established style, without me hand-holding every single prompt. I already had a rough first draft of instructions. What I didn't expect was how much back-and-forth it would take between two different AI tools, each doing a job the other genuinely isn't built for, to get there.
Two Tools, Two Jobs
Gemini generates the images. That's simply not something Claude does — so for the actual pixels, there's no debate about which tool does the work. Claude, it turns out, is better suited to a different job entirely: writing and iterating the instructions that tell Gemini what to do, diagnosing why an output missed the mark, and holding onto the accumulated lessons across dozens of rounds so the instruction set actually improves instead of drifting. Think creative director and QA lead, not illustrator. I'd draft a brief, hand it to the Gem, generate an image, bring the result back to Claude, and we'd pick apart what worked, what didn't, and why — then revise the instructions and go again. That loop — generate, review, diagnose, revise, retest — is where most of the actual work happened. Here's how it played out.
Round One: A Reasonable-Sounding Guess That Was Wrong
My original draft was solid on structure but thin on brand — it didn't actually know what this blog looks like. Claude's first pass filled that in with an educated guess: moody, cinematic, desaturated concept art, no text baked into the image, a single strong visual metaphor per post. Reasonable-sounding. Also wrong. The test outputs came back competent but generic — and worse, they'd all converged on the exact same template: a dim room, a blurred blue window, a warm spotlight, a dark tabletop. Different prompts, same photo shoot. “I don't think the instruction is working,” I said. That turned out to be an understatement.
The Reframe: Show, Don't Guess
Rather than keep refining a guess, I pulled seven real headline images straight from the published blog and handed them over as ground truth. That changed everything. It turned out my actual house style has almost nothing to do with moody cinematic art — it's bold headline text baked directly into the image, multi-panel infographics, before/after comparisons, flat icon-and-headline posters, and photoreal scenes with UI-mockup overlays. Informative and a little punchy, not atmospheric. Claude rewrote the instructions from scratch around that real archive instead of a first impression — five named format templates, typography rules for baked-in headline text, and a brighter, more confident palette in place of the desaturated cinema look.
Round Two: Right Content, Wrong Craft
The next batch nailed the content and the structure — but looked like clip art. Stock padlock icons, gavels, uniform drop-shadow boxes in a neat grid, a rainbow of primary colours all shouting at once. Technically on-brief, visually cheap. So we set an explicit reference bar — Bloomberg Businessweek, Harvard Business Review's digital cover art — and named the exact tells to avoid: generic icon-pack glyphs, mixed rendering styles in one image, evenly-sized boxes with identical shadows arranged in a rigid grid. The next round was a visible step up: real depth, tighter two-and-one-neutral colour palettes, asymmetric layouts instead of PowerPoint-style symmetry.
Catching the Bugs No One Would Have Guessed
This is where having a second tool doing the reviewing earned its keep. A few issues turned up that weren't obvious from a single image in isolation — only from comparing outputs across a run: Session anchoring. One generation, run straight after another in the same Gemini session, quietly reused the previous image's entire layout and content — different headline pasted on top of the wrong diagram. Running the same brief in a fresh window fixed it instantly, which told us it was a session-context issue rather than a flaw in the instructions themselves. A dropped rule. A cliché-avoidance list from an early draft — no stock handshakes, no glowing brains, no robot hands — quietly didn't make it into a later rewrite. A literal stock handshake reappeared two rounds later as proof. Comparing the current instructions against the original catch list is what caught it, and it went straight back in, with a few more retired clichés added for good measure. A production detail nobody would think to test for. I apply a rounded-corner crop to these images in HTML once they're published. A watermark placed flush in the true corner gets clipped by that radius — so the instructions now pull it in a few percent from the edge instead.
Where It Landed
After roughly a dozen test-and-revise cycles, the Gem now reliably produces images that match the actual blog archive: the right format for the content, real headline text, a consistent watermark, layouts designed to survive a centre-crop on social platforms, and — critically — a craft level that doesn't fall back to clip art the moment my back is turned. None of that came from one clever prompt. It came from a loop: generate with Gemini, diagnose with Claude, ground every revision in real evidence rather than assumption, and specifically go looking for regressions rather than assuming an improvement in one area didn't quietly break something else.
The Actual Takeaway
This isn't really an image-generation story. It's a process story, and one I'd argue applies well beyond blog graphics: the first output from any AI tool — or any new hire, playbook, or sales process, for that matter — is a first draft, not a finished product. The value showed up in the testing loop, in treating each round as evidence rather than a verdict, and in deliberately checking that a fix in one place hadn't quietly broken something you'd already solved.
Pick the right tool for each job, and don't assume the first pass is the last one. Occasionally right about AI, remember.
