Conversations that convert

arOS helps small businesses and agencies create marketing with AI. Its edge is context: reusable profiles of a business's brand, audience, product, and offer, which arOS calls DNAs. These profiles provide reusable business context for AI-generated responses.

When I joined in June 2025, arOS 1.0 worked agent by agent. You picked a specialist, such as an offer generator or a carousel writer, filled in its fields, and got one output at a time. From June to December 2025, I worked on 1.0 with a team of 6 low-code developers and 2 prompt engineers.

• Automatic DNA creation. A guided questionnaire replaced 40 to 60 minutes of manual forms. The agents then generated the profile in 1 to 2 minutes.
• Combinable DNAs. Three types, personality, audience, and product or service, that mix and match, so a new product or audience no longer meant starting over.
• Projects. Pick and preset several agents on one screen, instead of opening them one at a time.
• Agent info. In research, 5 of 10 users couldn't tell what an agent did from its name, and 7 of 10 said their first generation failed because of it. Every agent got a panel explaining when to use it.

Generic AI tools write fast. In our discovery interviews, people described a different bottleneck: refining. They spent their time fixing tone, swapping words they would never use, and reshaping sentences that didn't sound like them, often across several AI tools for a single post.

My PRD put it in one line: AI was speeding up writing, not results.

At the same time, more than 70% of the users we researched wanted to work through chat. For most people, generative AI already meant chat. And with 32K+ registered users on 1.0, the new version had to feel like an upgrade, not a reset.

How we would know it worked

• Time to the first answer the user rates as good, and onboarding activation
• Fewer manual adjustments to answers
• Answer quality scored with LLM-as-a-judge evals
• Retention by cohort, and cost and latency per answer

From January to October 2026, I was the designer and Product Owner of arOS 2.0, working with our Head of Product, 5 developers, and 2 prompt engineers.

• Wrote the 2.0 PRD: a lean first release, decision principles, the split of roles between Product, AI Engineering, and Technology, AI governance in Langfuse, and success metrics.
• Wrote the August 2026 product plan: a diagnosis from production data, a benchmark of 28 companies, audience research, and each initiative tied to a metric for evaluating its success, plus what not to do yet.
• Framed every new initiative with a template based on Shape Up pitches, and wrote the user stories, from simple features to new products and internal platforms for analytics, usage, and access management, with acceptance criteria and design specs tied to the coded design system.
• Structured the research scripts, co-led the interviews, ran usability tests, and reviewed PostHog every week.
• Designed the base UI for chat and DNAs, created and maintained the design system, and prototyped features in React before development, such as visual identity in the DNA.
• Built features myself with an AI-assisted workflow, such as the reply chips, after recordings showed how many messages were just "go ahead", and commands and shortcuts. Small, evidence-backed improvements reached users without taking time from a small engineering team.

Constraints

• 32K+ registered users on 1.0 to bring along.
• Credit-based usage: every message, edit, and generation costs credits.
• A small team, so uncertain bets had to stay cheap.
• Non-deterministic output: answer quality had to be measured, not assumed.

The interviews and usability tests shaped the base of 2.0: onboarding, the chat itself, how people choose skills and get results, and business context through DNAs.

Onboarding was the hardest part to get right. It went through three versions, each one driven by where people were dropping off.

The first version asked for links, then played a short animated tour of the chat, skills, and DNAs, which helped people coming from 1.0 find their way. A 5-minute welcome video followed, with a loading bar underneath showing the DNA being created from the links. When it finished, a button let people into the product.

To keep people doing instead of watching, we removed the tour and the video. After the links, a chat that already knew the business from them built the first DNA together with the user, through a few questions with quick-pick answers.

Once posts and Instagram content became a core job, Instagram joined onboarding: reading the profile helps us understand the business, and its palette, post style, and type feed the brand DNA. The links step also gained a context panel that explains what we need from people and why.

The first skill picker was narrow, with search on top and one long list. Skills with options expanded inside the list and stretched the popover. It hid what we had built: of about 100 skills, a long tail of around 75 averaged 3.4 selections a month each.

What changed, and why:

• Tabs by category, to help users scan skills more quickly by category.
• A grid instead of a list, showing more skills at once.
• Tags on each card, showing which areas a skill touches, such as content, strategy, ads, or copy, the same areas as the tabs.
• Search on the right, to give the tabs room.
• Options that open in place, without resizing the popover.
• The first option preselected, so a required option is never forgotten.
• Remove from the chip: a selected skill comes off right from the composer, without opening the picker.

After the first release, PostHog and the 550+ session recordings I analyzed shaped a second wave. We hadn't nailed our ideal customer yet, so we got closer to SaaS founders and saw their need to build credibility on Instagram. Most of what follows came from that.

People liked coming back, but creating was an event, not a routine, and generic output was a frequent complaint. That set the agenda for what follows.

Creating and updating DNAs without leaving the chat was a request since 2.0 launched. arOS infers the change from the conversation, shows its progress field by field, and links the result. It uses credits like any other message.

Posts and carousels weren't in the first release. They came when SaaS founders became our focus: carousels generated in the chat, an editor that takes both chat instructions and manual edits, Instagram connection and analysis, and publishing to Instagram, approved by Meta.

DNAs were organized in tabs, brand, audience, product or service, and offer, and every one was a text field. There was no good way to define the look of a post. People complained that posts didn't match their brand and spent a long time fixing them. Some gave up on the platform, others kept only the copy.

I watched the recordings, saw the frustration, and measured edits and regenerations: 13,722 cover regenerations in a single month. The post creator needed to talk to the user's DNA. Visual identity became a new section in the brand DNA, with post style, colors, type, ready-made templates, and a live preview. People can connect a different Instagram account at any time, and its palette, post style, and typography update the visual identity, so posts stay close to the brand they already have. During this period, arOS generated 20 posts per day.

The recordings showed people referencing other conversations, pasting screenshots of past chats, and trying to bring in documents. Memories let them save something once and have it apply to every chat, with a page to manage them: when the chat suggests a memory, 70% get accepted (36 of 51). Attachments followed, kept simple, and brought images into the chat, another frequent request.

Only 57% of users chose a skill when chatting. People liked the product but missed most of what it could do for them. So skills started finding the user: automatic skill inference picks the right one from the conversation, and commands and shortcuts, which I planned and implemented, let people call a skill the way they already did in other AI tools.

About 70% of visits came from phones, 7,046 against 2,988 on desktop, and cancellation feedback singled out the mobile experience. I redesigned the experience for small screens around one idea: the main actions within thumb reach. We shipped it as a PWA.

In August 2026, I turned this diagnosis into the product plan. It set the next bets:

• Production as a routine. With 80% of people producing on a single day of the month, the next step was a whole week of posts in one batch, copy first and visuals after approval.
• Closing the loop. Publishing to Instagram and LinkedIn, so results come back into the product.

I created and maintained the arOS design system: tokens in Tokens Studio, with a base set and semantic sets for light and dark, mirrored in code. When code became our single source of truth, it lived in Storybook, built on Tailwind and Radix.