arOS · Prompt engineering · 2025 to 2026
80+ agents, one brand voice
Models judged by cost per approved answer, not cost per token.
Marketing knowledge turned into prompts, a main chat that orchestrates the product, and a way to measure which answers are actually good.
Fact sheet
- Role: Prompt Engineer and Marketing Strategist
- Period: Mar 2025 to Sep 2026
- Team: developers and founders at arOS
- Scope: agent prompts, main chat, model evaluation, workflow maintenance
- Key result: 80+ agents designed and evaluated
Context
Every arOS agent started from the customer's DNA: brand, positioning, audience and voice. My job was the marketing logic inside the prompts and judging whether the output would actually work as marketing.
The problem
A well-written answer can still be bad marketing. Users would get fluent text that missed the offer, the stage of awareness or the brand's voice. And every model call has a cost.
What I did
- Designed prompts for 80+ agents. Strategy, storytelling, copywriting, VSLs, scripts, ideas, newsletters, carousels and an SEO, GEO and AEO agent that audits a page and recommends fixes.
- Owned the main chat. The chat that orchestrates the product adapts to the user's tone and company DNA, with a system prompt chained to skills, prompts, tools and workflows. I worked on caching and on protection against prompt injection.
- Measured quality and cost. In Langfuse I versioned prompts, read traces and evaluated answers. Through OpenRouter I compared GPT, Claude and Gemini families on quality, instruction following, brand voice, speed, tokens and cost. The benchmark used cost per approved answer, not cost per token.
- Built the flows behind the agents. arOS v1 ran on Bubble with n8n. I implemented the workflows, automations and RAG that feed the prompts, and worked on content import and transcription.
- Joined Iman Gadzhi's AI team. I was also part of Iman Gadzhi's AI team.
- Worked on the engine, not only the words. In 2.0 I worked on skill orchestration in the main chat and on the carousel engine, from the prompt to the code, and took part in product ideation.
Key decisions
| Evidence | Decision | Trade-off |
|---|---|---|
| Fluent answers were failing as marketing | Judge outputs against offer, audience and goal, not only writing | Slower evaluation |
| The cheapest token was not the cheapest result | Rank models by cost per approved answer | Needs repeated runs and human review |
| A timestamp at the top of the prompt broke caching | Move dynamic data to the end of the prompt | Had to revisit a fix already shipped |
Results
80+ agents built and evaluated with the team. The cache came back after the timestamp fix. We built the product as a team, and I worked across almost every layer: marketing prompts, flows, automations, workflows, RAG, skill orchestration, the carousel engine and product ideas.
What I learned
- Prompts are marketing documents. Most quality problems were strategy problems.
- Before shipping a fix, check what it does to quality, cost and speed across the system.
From 1.0 to 2.0
From 80 agents to one conversation.
arOS changed shape while I was there. In both versions I worked on the whole chain, from the product idea and the marketing prompts to the flows and code behind them.
Where I worked, end to end
- Product ideation
- Marketing prompts
- Flows and automations
- Workflows
- RAG
- Skill orchestration
- Carousel engine, prompt to code
- Landing pages in code
-
arOS 1.0
One agent per task
- Pick a specialist: strategy, copy, VSL, carousel, newsletter or SEO
- Select the brand DNA and fill in the request
- Edit the result, then move on to the next agent
My part The prompts of 80+ agents and the n8n and Bubble flows behind them: workflows, automations and RAG.
arOS 1.0: the Carousel Generator agent, with a personal brand DNA -
arOS 2.0
One chat that orchestrates
- Ask in the conversation
- The chat brings in the brand DNA and calls the right skill, prompt or tool
- Answers, carousels and posts come back in the same thread
My part The main chat’s system prompt and skill orchestration, the carousel engine from prompt to code, caching, prompt injection protection and the model benchmark.
arOS 2.0: the same request in the chat, with a real estate agency’s DNA
Same request, with and without the brand DNA
From the chat to a finished post
I worked on the carousel engine behind this, from the prompts to the code. The interface is Vitor Rocha’s design.