# I Built a Hermes Agent That Runs a Two-Post-a-Day AI Media Workflow > Source: > Published: 2026-05-20 05:10:08+00:00 This is a submission for the Hermes Agent Challenge. I built a Hermes-powered content operations system for a Chinese WeChat Official Account focused on AI tools, cross-border monetization, content going global, and practical AI workflows. The goal was not to make a generic chatbot. I wanted Hermes to do the unglamorous operational work behind a real media account: The account publishes one article at 7:00 AM and another at 6:00 PM. The morning slot favors practical tutorials and tool/workflow guides. The evening slot favors trend analysis, controversy, reviews, and case breakdowns. I also added a second operational workflow: a daily Xiaohongshu research queue for pet-supplies affiliate videos. It prepares search targets and review files, but intentionally keeps the final approval human-in-the-loop because the acceptance criteria require audio and on-screen text judgment. The WeChat pipeline produces a complete local article package for each run: ~/.hermes/wechat-mp/YYYY-MM-DD/morning/ article.md article.html article.json topic_research.md image_plan.md cover.png cover_prompt.txt The draft creation step uses the official WeChat API. A run only counts as successful if the API returns a real draft media ID. Otherwise, the generated package stays on disk for inspection and Hermes reports the failure. The Xiaohongshu workflow produces a daily review queue: ~/.hermes/xhs-pet-supplies-videos/YYYY-MM-DD/ review_queue.txt approved_links_template.txt Approved video links are appended to: ~/.hermes/xhs-pet-supplies-videos/links.txt The important design choice: Hermes automates the repeatable operational work, but it does not pretend to verify things it cannot safely verify. For videos, "no Chinese narration" and "no more than 10 Chinese characters on screen" need visual/audio review, so the system creates the queue and saves only manually approved links. The private local deployment has been tested end-to-end with real scheduled jobs and real WeChat draft creation. I am not publishing account credentials, access tokens, draft media IDs, or screenshots that expose private account metadata, but the workflow is designed around concrete success checks: article.md , article.html , article.json , topic_research.md , image_plan.md , and cover.png .Repository: github.com/kax168/hermes-agent-content-ops The private installation contains: Secrets are stored outside the repo in ~/.hermes/.env and are never printed or committed. Hermes is the operating layer of the project, not just a model wrapper. I created a local Hermes skill called wechat-official-account-operator . It contains the account strategy, content rules, visual requirements, API safety defaults, output schema, and success criteria. That mattered because the account has a strong editorial position: The skill keeps these rules close to the agent so each run is not starting from a blank prompt. Hermes cron runs two jobs every day: 0 7 * * * for the morning article;0 18 * * * for the evening article.This is where Hermes felt different from a normal chat interface. The system is not waiting for a human to ask, "Please write a post now." It wakes up, collects context, generates artifacts, calls an API, and leaves a concrete result. The WeChat pipeline performs several real steps: article.json .I added a strict "no simulation" rule: the run is not successful unless the local files exist and the WeChat API returns a real draft result. One of the more interesting lessons was that a fully autonomous large-prompt agent run was not the most reliable way to handle scheduled publishing. Some agentic runs timed out or got stuck on browser/model calls. So I changed the architecture: Hermes still owns the schedule, skills, memory, and operational workflow, but the daily execution path uses deterministic Python scripts for the fragile parts. That hybrid design made the system more useful: The Xiaohongshu workflow is deliberately semi-automatic. It searches for pet-supplies affiliate video candidates and prepares a review queue, but final saving requires manual confirmation. That is not a limitation I wanted to hide. It is a product decision. If the rule says "no Chinese narration" and "almost no Chinese on-screen text," then silently adding links without checking would produce bad data. Hermes is still useful because it removes the repetitive search/setup work while preserving the judgment step. The Build With Hermes Agent prompt is judged on effective use of Hermes Agent's agentic capabilities, technical implementation and code quality, creativity and originality, and usability/user experience. Here is how I designed for those points: This project is small, but it is real. It has all the annoying parts of actual automation: Hermes Agent was a good fit because the workflow needed memory, scheduling, tools, and persistence more than it needed a prettier chat response. The result is an agentic system that can operate a narrow but real content pipeline: from topic discovery to WeChat draft creation, twice a day. I would like to add: The biggest takeaway: the best agent projects are often not magic demos. They are workflows where the agent keeps showing up, on schedule, doing the boring parts correctly.