One URL, Six AI Calls: How We Turn a Website Into a Content Calendar
The Embarrassing Pitch
Here's what we tell people at demos: paste your website URL, and in about ten minutes you'll have two weeks of Instagram-ready content featuring AI influencers holding your real products, complete with captions, hashtags, and scheduling.
It sounds absurd. Most founders who hear it assume we're faking it. We're not. But making it real required solving a harder problem than the pitch suggests — not the image generation or the AI writing, but the coordination. How do you get a chain of AI calls to produce something coherent when each one knows nothing about what came before?
The Waterfall Problem
Our first instinct was sequential: scrape the website, analyze the brand, pick influencers, plan the calendar, generate content. Simple enough on paper.
The problem is that sequential AI pipelines don't stay simple. Each step has context it needs from the previous step, but also context the previous step didn't know it would need. The influencer matcher needs to know what the brand sounds like. The calendar planner needs to know which products exist and which influencer is assigned. The caption writer needs the calendar slot, the product details, the influencer's voice, and the brand's tone — all at once.
By the time we mapped the actual dependency graph, we had six distinct Claude calls, plus a parallel image generation layer on top. The question wasn't whether to use AI — it was how to structure information flow so that each call got exactly what it needed without re-doing work the previous step already did.
Intelligence at Every Layer
The first call does the heaviest lifting. We feed it everything we can find about a business — website text, product images, Instagram post history, bio, engagement signals — and ask it to produce a structured brand profile. This is the document every other call reads from.
What makes this work is that the brand profile isn't just marketing copy. It includes product-level data extracted from real Instagram captions, color palette signals pulled from CSS and meta tags, visual mood descriptions derived from the existing feed, and audience inference from the follower and engagement patterns we can see. It's closer to a competitive intelligence report than a brand brief.
This profile feeds everything downstream. The theme analyzer uses it to distill a coherent visual direction. The mood board layer translates that into concrete image aesthetics. The influencer matcher scores each AI persona against the brand's industry, tone, and visual style — not as a lookup, but as a reasoned match with explanation.
Anti-Hallucination by Design
One issue surfaces quickly when you ask AI to extract products from a website: hallucination. Claude would occasionally invent product names that sounded plausible but weren't real — synthesizing what it thought the brand probably sold based on industry patterns.
We solved this with a validation layer between the brand analysis call and everything downstream. Every product name Claude returns gets checked against the actual source corpus: the website text, Instagram captions, and image alt text we scraped. If fewer than half the significant words in a product name appear in the source material, that product is dropped. It's a simple word-intersection heuristic, but it catches the vast majority of hallucinated products before they propagate into image generation and captions.
If filtering removes everything — which happens occasionally with very sparse websites — we fall back to the full list. A calendar with potentially hallucinated products is better than no calendar at all.
The Calendar as Contract
The calendar generation step is where the pipeline's structure really matters. This call produces the scheduling skeleton: which content type each slot will be, which product it features, which influencer is assigned, what angle the caption should take. Every call after this one reads from the calendar rather than re-reasoning about these questions.
This matters for consistency. If you generate captions and images independently and only connect them at the end, you get a portfolio that doesn't cohere — the image might show a product in a casual outdoor scene while the caption is selling the same product with formal business language. By treating the calendar as a shared contract that all downstream calls read from, we keep both channels aligned on intent before either one does its work.
Parallel at the Right Level
Once the calendar exists, we parallelize aggressively. All six posts generate their captions simultaneously. All six posts generate their images simultaneously. Quality scoring runs across all posts together at the end.
What we don't parallelize is the planning layer. Brand analysis, theme, mood board, influencer matching, and calendar generation run in a structured sequence — each one informed by the output of the previous. Trying to parallelize this phase saves a few seconds and costs you coherence.
The image generation phase is where the economics get interesting. A single UGC post — an influencer holding a product in a realistic scene — can require two to three separate model calls: one for the influencer rendering, one for the product composite, one for the final scene integration. A six-post calendar might involve eighteen image generation calls in total, all running in parallel.
One Pipeline, One Answer
The result of all this coordination is that a business owner with a working website URL gets a complete, opinionated content plan in a single operation. They don't configure anything beyond the URL. They don't choose influencers or pick content types or brief a designer.
The pipeline reasons about their brand, makes choices, and delivers an answer. Sometimes the answer needs adjustment — they can regenerate individual posts or swap influencers — but the default output is good enough to publish.
That's the real design goal: not just automating content creation, but automating the judgment calls that make content creation expensive. Figuring out what to post, who should post it, how it should look, and what it should say — that's the work we're trying to eliminate, one URL at a time.
Jiwa AI is built for Southeast Asian small businesses who want social media presence without the agency retainer. If you're curious about the pipeline, try it on your own website.