Research, decompose, create, measure, refine

For the first few months of building Naomi, she worked in a creative vacuum. We'd tell her what to make, she'd make it, and the quality depended entirely on how good our brief was. She had no independent understanding of what was performing well on the platform, what hooks were trending, or what visual styles were resonating with audiences.

That's like hiring a social media manager and never letting them scroll TikTok.

The research layer

The first piece was connecting Naomi to live platform data. sync_tiktok_profile_posts pulls up to 200 posts from any TikTok profile via Apify, with full engagement metrics — views, likes, comments, shares, saves. research_hashtag surfaces top-performing videos for a given hashtag. These aren't static databases; they're live queries against the platform.

When Naomi starts a content session now, she can look at what's trending before she creates anything. She sees which hooks are working, which formats are getting engagement, which topics are resonating. This isn't revolutionary — every human social media manager does this. But for an autonomous agent, it's the difference between informed creation and random generation.

The decomposition engine

Knowing that a video is performing well isn't enough. Naomi needs to know why. That's where video decomposition comes in.

Given a TikTok URL, the system downloads the video and runs it through a four-stage pipeline: ffmpeg for scene detection, Gemini Pro for shot boundary refinement, Flash workers for per-shot analysis, and a Pro synthesizer for narrative structure. The output is a structured blueprint: what each shot contains, how the camera moves, where text overlays appear, what the hook-build-payoff structure is, and — crucially — a recreation brief that explains the creative DNA well enough to make something inspired by it.

The decomposition isn't copying. It's pattern extraction. "This hook works because it opens on a close-up of an emotional expression before revealing context" is a transferable insight. Naomi can apply that pattern to her own character, her own message, her own visual style.

The research UI

This is where it gets interesting for humans too. The research tab in Naomi's workspace shows decomposition reports in a structured viewer — shot by shot, with the original clips, metadata, and analysis. Each section of the report (hook, build, payoff, visual style, recreation brief) has an "Ask Naomi" button that threads the context directly into agent chat.

So you can be reading through a decomposition, see something interesting about the hook structure, click "Ask Naomi about this," and she gets the full context of what you're looking at. No copy-pasting, no describing what you saw. The UI hands off context seamlessly.

The "Draft from report" button goes further — it sends the recreation brief to Naomi as a creation prompt, so she can start building content inspired by the research in one click.

The loop

Put together, the flywheel looks like this:

1. Research: Sync trending posts, identify what's working
2. Decompose: Break down top performers into structured blueprints
3. Create: Build new content using the patterns, with Naomi's character and voice
4. Measure: Track engagement via review_performance
5. Refine: Update account learnings with what's actually working

Each cycle makes the next one better. The account profile accumulates learnings: themes to try, themes explored, strategy notes. Session after session, Naomi's understanding of what works for this specific audience gets sharper.

The goal isn't to copy what's trending. It's to understand why things trend, and use that understanding to create something original that benefits from the same structural insights.