Huebloom

I’ve been building Huebloom, a website that offers free printable coloring pages and PDF downloads. The site now contains a large and growing number of pages, but I don’t create them manually one by one. Most content websites follow predictable structures: every collection needs images, titles, descriptions, downloadable files, and the same page components. That made Huebloom a good candidate for an automated production pipeline. ## Why automate it? The work is highly repetitive. Once the structure and quality standards are defined, manually rebuilding the same page format adds very little value. I wanted AI agents to handle the repeatable parts while I focused on the decisions that shape the product. ## The pipeline: from keyword to published page ### 1. Monitoring search demand An agent monitors Google Trends and other keyword signals for terms that are rising quickly but already have some baseline search demand. My working assumption is that a sudden increase in demand can create a temporary content gap. Search engines need more relevant results, which may give newer websites an opportunity to compete. It is not a guarantee of rankings, but it is a useful signal for deciding what to investigate. ### 2. Research and planning When the system identifies a promising term, AI researches the topic and starts a brainstorming process. It looks at the available information, related searches, existing pages, and possible subtopics. From there, it plans which collections and supporting pages could be useful instead of generating a single isolated page. ### 3. Generating the images The image workflow uses existing visual references as inspiration, then generates new compositions in batches. This stage also includes processing the images, checking formats, preparing printable versions, and organizing the assets so they can be connected to the correct collection. ### 4. Producing and publishing the content AI supports the rest of the workflow: naming assets, drafting titles and descriptions, preparing page data, creating downloadable PDFs, checking the page copy, and publishing the finished collection. The mechanical pipeline is automated from discovery through release. That does not mean everything is accepted without review. ## Automation needs its own feedback loop The most useful part of the system is what happens after each repeated task. The agents review what went wrong, where time was wasted, and which requirements were unclear. When I find a recurring problem, I turn the correction into a reusable rule and record it in the system’s Skills or Memory. The automation still makes mistakes. The difference is that the same mistake should become less likely with each run. Over time, the workflow becomes faster because more of the judgment behind repetitive tasks has been made explicit. ## What I still keep human Three parts still require final human judgment: **Visual quality.** AI image generation is not completely predictable. Anatomical mistakes, awkward compositions, inconsistent lines, and small visual defects still need a human review. **Vibe and taste.** The overall visual direction of the site cannot come from automation alone. Someone still needs to decide whether the pages feel coherent, enjoyable, and appropriate for the audience. **Tone of voice.** AI can draft descriptions, but the final language needs to sound like the same product across every page. I still review how the site speaks. The system automates execution, but it does not replace judgment. That distinction is what makes the workflow useful for me. A single person can build and maintain a content-heavy website without manually repeating the same tasks hundreds of times, while still controlling the parts that shape the identity and quality of the product. This is the project: https://huebloom.art/ For anyone else building agent-based workflows: which parts have you successfully automated, and which decisions do you still keep human?