project 02 / 09

Nine agents, one social media team

Designing a team of AI agents that fully understands the brand before starting any project.

AI Agent DesignMulti-Agent OrchestrationPrompt EngineeringContent StrategyBrand Voice SystemsUX WritingCreative DirectionWorkflow DesignClaude CodeGeminiNano Banana Pro

challenge

Solo creators and lean teams make social content without the thing that actually makes it good: other people in the room. No strategist to pressure-test the plan, no creative director to push the concept, no editor to catch the tone. AI can stand in for some of that, but most AI content workflows skip the part that matters and go straight to generating posts, so the output is fluent and generic. I wanted a system that worked the way a real agency does: understand the brand first, then plan, concept, write, and produce, with a human approving each step.

my role

Designer and author of the whole system. I defined the roles, the order they work in, the rules each one follows, and the hand-offs between them. I also wrote the separate methodology they depend on, a brand DNA extraction process that has to run before any creative work starts.

methodology

01Roles before prompts: mapped the functions a small agency performs and wrote each as its own agent with a single job. Orchestrator for intake and gap-checking, Planner for the content plan, Trend Researcher for references, Creative Director for the big idea and key visual, Writer for copy and alt text, Social Media and Editor for production, Content Multiplier for adapting finished work to other formats.

02DNA as a gate: nothing gets planned or designed until a brand DNA exists. A preliminary agent runs the extraction (positioning, audience, voice, visual system) and the rest of the team reads it in full before starting. A brand without one is a hard stop.

03One question at a time: every agent asks a single question per message and assumes nothing. Approvals are explicit and piece by piece, and the visual direction has to be signed off before any final asset is produced.

04Two engines, one team: the same system runs in two environments. In one it generates the images directly, in the other it stops at ready-to-use prompts. Language mirrors the client automatically, English or Portuguese, without announcing the switch, while file names and image prompts stay fixed.

05Tested on real briefs: ran the full workflow on three brands with different constraints, a celebrity beauty line, a Portuguese-language workshop launch of my own, and a seasonal campaign for a matcha brand.

outcomes

→A working team of nine agents portable across two AI environments (Gemini and Claude).

→Turned "understand the brand" from a good intention into a required step the workflow cannot skip.

→Made the parts of my own creative process explicit instead of leaving them in my head: the editorial filter, the approval checkpoints, the visual consistency rules.

→Confirmed the premise across three very different briefs. With real brand context loaded first, the agents give a solo creator the second, third, and fourth opinion they don't otherwise have.

→Documented the whole system so it can be handed to someone else and run without me.