Case study · Orita

Design is the stage, content is the performer

Orita’s brand grew up in two years. Research changed what it had to do, every surface got rebuilt on one set of rules, and then I wrote the rules down so a team of non-designers and a machine could hold them without me.

Role: Sole designer; contractor, continuous since 2021 and ongoing
Shipped: Orita brand refresh, May 2026Slide design skill, May 2026Claude skills workshop, May 2026Org-default Claude design system, July 2026orita.ai/brand, July 2026
Timeline: 2024–2026
Site: orita.ai

2024

Where the brand came from

Until 2024, Orita’s brand was a color: yellow. That year the CEO wanted the messaging and the brand made more polished, something that would land with the customers we were selling to, and I worked with him to get there. I wrote a research brief, and the week of May 13 I ran the team through a positioning workshop: personality, value proposition, messaging pillars, a per-persona messaging strategy. It gave us our first messaging strategy, and the five personality traits that still define the brand today: dedicated to finding the best solution, smart and analytical, friendly and approachable, humble and chill, empathetic and reassuring.

I was on product at the time, so the CEO, wanting to move fast, hired another designer for the site and the logo. Not fully satisfied with where it landed, he brought me in, after the logo, palette, type, and visual style were decided, to take about a week and clean things up, with a fifteen-minute overview of the thinking behind it. So I did what you do with a system you didn’t design and can’t re-litigate: I gave it rules. Semantic definitions for every shade, a use case for every style and pattern, master styles tied across the web app and the decks. In that first week I redesigned seven pages and drafted the team’s initial brand guidelines.

The outside designer's V1 palette proposal: a wide yellow band over coral, peach, sky blue, pale blue, and cream swatches, a dark charcoal, and greys; below it an expanded set of rounded swatches with a peach-to-blue gradient. — **The palette I was handed** The outside designer’s proposal: a yellow, a charcoal, a coral, a blue, and no jobs for any of them. My first week gave every shade a semantic definition and every style a use case.

The 2024 homepage as handed over: a yellow announcement bar, the headline Land in the inbox, grow profitably, a yellow Make more sales button, and a hero photograph of cookware with floating stat cards. — **The homepage as handed over, August 2024** Built by the outside designer; mine to tidy in a week. The playful register was set here, and I ran with it.

2024–2025

A brand that grew playful, and a team that couldn’t hold it

That system fed straight into the decks, where most of Orita’s brand shows up in the wild. The team was ramping up pitches and partnerships, building decks from scratch each time, and most of them ran through me for polish. So I turned the slides we reused most into a canonical set the team could pull from.

Through 2024 the direction from leadership was more personality: playful illustrations built around Atiro, the yellow octopus. I leaned into “friendly and approachable,” and the brand grew bright. The logic was reasonable: visual effort put distance between us and the WordPress-and-Google-Slides look the team had previously been living with.

Canonical slides anchored the spine of our decks, but customer success isn’t a design team, and the bespoke slides they added looked amateur next to the rest of the deck. It risked exactly the perception the rebrand was meant to fix. So in late 2024 I specced a modular system, a wider set of templates plus a component library to copy and rearrange. By July 2025 it existed as team-wide Slide Templates with a recorded tutorial.

The July 2025 slide templates The playful brand at its brightest, in my own hands

A team-built slide titled Engagement-based Best Practices, continued: a wall of small body text with the labels Email 5, Email 6, and Post-Purchase Follow-Up Campaign each highlighted in yellow, green check emoji and a warning emoji as bullets, and no single point to land on. — **What drift looked like: everything important highlighted in yellow** A team-built slide from the era. The body text carries the whole point, the highlights compete with each other, and the emoji stand in for a hierarchy. None of it the author’s fault; the system asked them to make design decisions it should have made for them.

A team-built results slide, Rescued profiles: Clicked to Order days: a bordered table with two rows filled solid green for the totals, beside a column of explanatory text. — **What drift looked like: decoration doing the hierarchy’s job** Solid green fills to mark the totals, a bordered grid, and a wall of explanation beside it. The same data, neutralized, is the after slide further down the page.

Late 2025 – February 2026

Listening

In the second half of 2025 I ran a research program nobody had assigned me: interviews with customers, agency partners, and the internal team, alongside a dashboard audit, a card sort, and a competitive teardown. I wanted to know how customers actually thought about Orita, what our surfaces were doing for them and failing to do, and what the product should change first. The answers came back about the brand as much as the product.

In February the co-CEOs asked me to lead an all-hands brainstorm on retention. I built the board, ran the session, and synthesized about a hundred and fifty stickies into five problem statements for leadership. The first was the one the whole team had circled without naming: customers can’t independently see or prove Orita is working, so the value story lives in manually assembled slide decks.

My own stickies on that board were the argument I’d make a month later: an agency partner keeping clients out of our dashboard because they wouldn’t understand it; a product UI that wasn’t set up to sell itself; core concepts explained nowhere but in the presentations we gave. It’s hard to build trust in something you can’t understand.

The picture that came back was consistent. Customers didn’t fully trust how Orita worked because they didn’t understand it. They weren’t sure how to use our segments alongside their own. Getting to value took longer than expected, and the product didn’t always match what the sales process had promised.

Five things converged

We marketed ourselves as fast and easy, and the research said Orita was neither for the customers living with it. The exciting yellows, the cute illustrations, the green arrows on every win: none of it matched their experience, and the customers we wanted to win expect an experience as serious as their list sizes. If you have a billion-dollar list, who would you trust with it?

The templates had helped, and drift continued, because the team didn’t have the bandwidth to learn a system with that many rules. The template library was outgrowing its users; more templates meant more scrolling. Everything routed through one designer, and some of the drift across the site, the decks, the portal, and the print pieces was mine. And the company had just put everyone on Claude, with skills newly available. Rather than keep optimizing a template system that wasn’t scaling, I started asking whether a written system could do the work better than a template library ever would.

March 2026

The audit, and the argument

I audited every surface we had, to find the common failures and to ask what a more intentional system would look like: one that met the company’s goals, fit the traits, matched what customers actually experienced, and could be reproduced by people who aren’t designers. The argument I took to leadership was that the brand and the product were selling different things: the brand was bright and playful, the product was sold on expert-led science. They agreed to publish in April.

The operating principle, in the words I used with the team: design is the stage, content is the performer. The first thing a customer reads should be the content, not the system holding it. Personality moved from the visuals into the copy. The rules under it are the ones I argued to the team, slide by slide.

A slide from the rationale deck titled Yellow is punctuation, not a surface: text explaining that yellow earns its place at key conversion moments and is never a body text color, card fill, or divider tint, beside six slide thumbnails marked with green checks and red crosses, the crossed ones using yellow as headlines, fills, and full backgrounds. — **Yellow is punctuation, not a surface** One moment per section, at most. Conversion moments and eyebrows only; never a body text color, never a card fill, never a divider tint. When yellow is everywhere, the moments that deserve attention lose it.

A slide from the rationale deck titled Data is neutralized: text explaining that stat displays are neutralized in the spirit of humble and chill, with no colorful accents, up arrows, highlighted rows, or card containers, beside three example slides showing plain stats, a plain table, and an engagement-tier table with a single bolded value. — **Data is neutralized** No arrows, no highlighted rows, no card containers around numbers. Bold the one value that matters; green text only on a revenue-growth value. Let customers get excited about the data when they understand it, not because we’re forcing it.

Color variation comes from the customer’s own content, so the per-customer pastels retired. Illustration went true-to-life: real UI, real data, clean diagrams. Four visual registers, marketing, teaching, conversion, celebration, so playfulness has a sanctioned home and never leaks into the product story.

April–May 2026

The homepage first

I started with the homepage: the gateway to the product and the brand, and the natural foundation for every other surface. The page I was replacing was a product showcase, not the calm, sharp partner the personality described. The redesign restructured it as a diagnosis: what it is, proof it works, how it works, then channels. Its centerpiece is a How It Works section that shows the actual data flow and the model, built as an animation, which became the brand’s illustrative reference point.

The early 2026 homepage hero: a yellow gradient field with a halftone texture, the headline AI Customer Segmentation to Maximize Campaign Revenue, sparkle icons, three sparkle-bulleted benefits, and a yellow Get a Demo button. — **Before, January 2026: the hero** Yellow as a background field, which shifts the register from precision to promotion. A label for a headline. Sparkles as bullets. The page read like an anxious salesperson rather than a guide.

The June 2026 homepage hero: a white ground, the headline Maximize Revenue with AI Customer Segmentation, three quiet benefit phrases, one yellow Get a Demo button, and three small green sparkline charts with the stats 26, 30, and 31 percent. — **After, June 2026: the hero** Yellow once, on the conversion moment. A claim for a headline. Three sparklines instead of three sparkles: proof before decoration.

The early 2026 How Orita works section: three yellow cards, the first with a cartoon octopus in a lab coat at a laboratory bench, the second a Klaviyo segments list, the third a dashboard screenshot, each with a numbered step beneath. — **Before: how Orita works** Three numbered steps and a cartoon octopus in a lab coat. Outcomes shown; the intelligence behind them never.

The June 2026 How Orita works section on a dark ground: a labelled diagram of inputs flowing into a model, three process bands for channel engagement, propensity to buy, and non-human activity, and outputs into a Klaviyo segment list from Highly Engaged to Not Engaged. — **After: how Orita works** The actual mechanism, animated: inputs, process, outputs, every subscriber ranked by engagement every day. This section became the brand’s illustrative reference point.

Two things I didn’t solve: the cool greys against the warmer off-white ground, deferred as a larger effort, and the co-CEO’s objection that the system was a suit with no pocket squares, which got a grace period rather than a design answer. The refresh shipped in May, and I carried its foundations across the rest of orita.ai and into the blog, where each header illustration is now a diagram of the mechanism the post explains.

Spring 2026

The slides, on the same principles

The decks had their own diagnosis: the templates weren’t being used because the team makes dense, multi-topic slides that templates can’t hold, and every from-scratch deck was a chance for the brand to drift, because the team was making styling decisions they shouldn’t have had to make.

A team-built slide titled What is Orita Grow: a bulleted definition on the left, a yellow atom-and-arrow illustration with the words Highest propensity to buy in the middle, and a dense column of small-type instructions with a product screenshot on the right. — **The pattern on a dense slide** Three content zones. Body text shrunk to fit the right-hand column. A heading that labels the slide instead of making a claim. A decorative illustration in the position the point should occupy. Nothing a reader can land on in five seconds.

So I simplified the entire system, to cut the number of decorative decisions anyone has to make. One decision was a deliberate walk-back: the original guidelines had banned Regular, and I brought it back because Bold and Regular are a single Cmd+B apart and customer success was losing time in the style dropdown. One concept per slide, with the dense version pushed to the appendix as the leave-behind. Some of it I lost: the paper texture I’d introduced wasn’t what I’d hoped for, and I switched to white.

A type specification slide: a 10pt Figtree Bold eyebrow in grey or yellow, a 22pt Figtree Bold heading, 14 to 18pt regular body in dark grey, a stat label with a large bold value sized from 24 to 120pt, and 12 to 14pt captions. — **The type spec** Figtree only. Headings 22pt Bold, claims of four to twelve words; body 14–18pt Regular; eyebrows 10pt and discouraged; stat values sized by count, from 120pt for one down to 32pt for five, never auto-shrunk.

A slide grid specification on a yellow ground: a fixed eyebrow band and a fixed header band at the top with their positions and sizes in inches, a large content zone labelled place content here, avoid zones marked above and below, and the Orita mark locked bottom-right. — **The grid, in inches** 0.75-inch gutters, a fixed eyebrow band, a fixed header band, a content zone, the logo locked bottom-right, and marked avoid zones. Google Slides’ default bullets, because fighting them is where drift starts.

The rules are checkable. A validator treats Semibold, Medium, Light, and Italic as hard errors, flags label-style headings, and warns at three or more content zones or about 75 body words. Writing the rules so a script can check them is what made them hold.

Then I designed for the tool as well as the team. I used the audit to find the layouts the team actually reaches for and designed those as canonical templates, simplified and reserved, the identity carried by hierarchy rather than decoration. Two master decks came out of it, canonical slides that get cloned and templates that get filled, which greatly simplified the logic anything downstream had to follow. Fewer decisions means both the team and the tool make fewer mistakes.

Nine of the twenty-nine slide templates laid out as a contact sheet: title, dark section divider, text content, two-column, customer case study with photo and three stats, single hero stat, simple table, stat-with-chart, and bullets beside an image, every text field a double-brace placeholder and the Orita mark locked bottom-right on each. — **Nine of the 29 templates** Every field a placeholder, so a human or a machine can tell at a glance what goes where. The identity carried by hierarchy, not decoration.

The grow-segment slide before the system: engagement tiers carried by solid green fill bands that fade down the table, three of the tiers sharing a single description, then grey bands for not engaged and for bots. — **One slide, before** Engagement tiers as solid fill bands

The same slide through the design system: the fill bands replaced by rows separated with hairlines, each tier marked by a small coloured dot, the two columns headed Segment and What it means, and the help-center line set on a footnote band above the Orita mark. — **The same slide, after** Tiers carried by dots on hairline rows

May–June 2026

Teaching a machine the brand

Then I wrote the taste down. I built orita-slide-design, a Claude skill that generates native, on-brand Google Slides from the canonical decks and templates, used it to rebuild our key decks, and put it through nineteen rounds in five weeks: generate mock decks from real customer-success use cases, critique, tighten the rules, generate again. A generated slide reached customer-facing use three days after v0.5.

The orita-slide-design skill's SKILL.md open at its When-not-to-use section: boundaries for brainstorm drafts, non-deck deliverables, non-Orita decks, legacy decks, the linked case-study library, and decks produced by external automated systems. — **Governance written into the skill** A pattern taxonomy, explicit “when not to use” boundaries, a rule that the case-study library is linked and never regenerated, and delivery into one shared folder so nothing got rebuilt twice.

What I learned, I wrote to the team as I released it: brand tokens, do’s and don’ts, and a pile of good and bad examples weren’t enough, and the biggest failure mode wasn’t the model. So I set expectations: the tool wasn’t there yet, other design teams were hitting the same wall, and the fix was slow documentation of rules we’d never written down.

The post to the team, June 2026 The uncomfortable part: a spec that reads complete to a human and leaves dozens of micro-decisions open. Every gap becomes a coin flip per generation, and the misses compound. Writing a brand spec for an AI is closer to writing a config schema than a creative brief.

A few weeks in, the team voted a skills session with me the top investment-day topic, and I ran it with the live demo built on a salesperson’s real prospect-research checklist. Same day, a teammate built her own skill; within a week a company-wide voice-of-customer skill existed. The skill itself was an okay answer to the branding problem and not a great one for the team’s range of decks: ask it for a layout it didn’t have and it found its way back to one it did.

The Investment Day Claude Skills deck open in Google Slides: the cover reads Claude Skills, turn the work you repeat into something you build once. The speaker note below describes the difference between asking an AI to do a task and teaching it to do that task your way, every time, so you never have to explain it again. — **The May 2026 workshop deck, itself skill-built** Turn the work you repeat into something you build once. The speaker note carries the thesis.

July–August 2026

The system that retired the skill

By then I’d concluded that training a model to produce on-brand work is writing instructions complete enough that a person with no context could execute the design, structured so a machine can follow them. So I ran a second audit, this time on the refreshed surfaces, and codified rules for every style, pattern, and component, designed for generative output from the start: principles the tool follows even when no template covers the request, and as much about when not to use something as when to, because the gaps are where drift lives. The brand became a parent tokens config plus an org-default Claude Design system, with coded layers alongside it, and the brand guidelines moved in; it’s where the brand lives now, because it’s where the team can reach it. Slide output lands roughly 80% of the way to the ceiling I could see, by my own five-criterion audit and as a conservative estimate; the remaining fifth is refinement I’d still make by eye, not output coming back off-brand.

A rules file open in Claude Design, orita-apply-slides.md, section 1, Principles: a prime directive to make one point land in about five seconds, then numbered principles, each with an Apply line and a Check line, covering one clear focal point, fitting the layout to the content, defaulting to less, scarce emphasis, silent design, and headings that claim while eyebrows label. — **One page of the system’s rules** Principles the tool follows even when no template covers the request, each with what to do and how to check it. The design is silent; the content speaks. Headings claim; eyebrows label.

When the system outperformed the skill, I retired the skill, on August 4, 2026. The goal was never the tool. It was a centralized source for the brand, reachable by anyone at Orita without me.

The Orita Design System project in Claude Design: the readme describing the paper and data-made-visible registers, and a status panel reading Published, Currently org default. — **The receipt** “Published · Currently org default” — every new project at Orita starts from this system

What I claim

ClaimSource

The system is published in Claude Design and set as Orita’s org defaultdocumented

I retired orita-slide-design on August 4, 2026, once the system outperformed itdocumented

Slide output lands roughly 80% of the way to the ceiling I could seerecollection

Whether it reduced the design work a deck takes, or kept the people who tried itnot measured

The numbers that would settle it: the share of decks that ship without a designer touching them, correction time per deck before and after the system, and how many people generate something in their first month against how many come back.

Documented rows have a dated artifact I can produce on request. Recollection is mine, undocumented.

The loop

In July 2026 an engineer stood up a working draft of a customer-facing trial-results page in a few days: fast, and rough the way a first pass is. I duplicated it into the design system and passed it through, spent a few hours refining the output by hand, then fed the refinements back into the system. I synthesized what customer success, customers, and leadership each needed; the engineer pulled pieces of it into his build. Rough draft to production-ready in under a week, the team’s own recap, and the infrastructure got better in the same pass.

Before — the working draft Two stacked tab rows, competing hierarchy, more options than the page needed · mock data throughout

The SMS Trial Results page after the design pass: a single aggregate view led by a plain-language headline about the best segment's return, with ROAS bars, a spend-versus-revenue chart, and a per-segment table. All data is mock data. — **After — through the design system, then by hand** Rough draft to production-ready in under a week · mock data

Working without me

The co-CEO generated his own on-brand deck; the file is literally named “Aaron made with claude.” Seven decks from five uploaders landed in the shared folder in its first six weeks. The marketing lead built a LinkedIn banner in the system. The brand became a public product: orita.ai/brand, a self-serve co-branding portal for partners and agencies. And at an August all-hands, unprompted, a teammate said the system’s output was “better than I could have come up with on my own, and it fit perfectly.”

The co-CEO’s own deck, June 2026 The filename is the caption · email redacted

orita.ai/brand, the public co-branding portal Live since July 2026 · the top of the page; lockups, colour tokens and type specimens run on below

What I claim, and what I don’t

Mine — the 2024 positioning workshop·systematizing the 2024 identity·the late-2025 research and the retention synthesis·the 2026 refresh argument, the homepage, and every page of orita.ai·the slide system, templates, and skill·the org-default design system·orita.ai/brand·the trial-results page design pass

Theirs — the logo and 2024 visual direction (two outside designers)·the five traits, written together in the workshop·the automation decks’ variable mechanism (an engineer) and copy (the co-CEO and customer success)·every deck the team generates inside the system·the skills teammates built after the workshop

Whether the system reduced the design work a deck takes, or kept the people who tried it, I never measured, and nothing I have access to says whether the refresh moved traffic or leads. The numbers that would settle the first: the share of decks that ship without a designer touching them, correction time per deck before and after, and how many people generate something in their first month against how many come back. Seven decks from five uploaders is the closest thing I have, and it counts one folder.

Reflection

For years my job was encoding other experts’ judgment into products that work when the expert isn’t in the room. This time the expert was me, and the hard parts weren’t the ones I expected. The research had to overrule my own pages; restraint is an easy principle to state and a harder one to apply to work with your name on it. And the judgment didn’t scale when I taught it or templated it. It scaled when I wrote it down at the level of detail a machine needs, which turned out to be the level a busy teammate needs too. That’s the job now: fewer decisions for everyone else, so the decisions left are the ones only a designer can make.