{
  "id": 7222776,
  "title": "From Claude Project to Hybrid AI Agent: Lessons from a Real-World Content Workflow",
  "url": "https://urgent.news/2026/09/14/from-claude-project-to-hybrid-ai-agent-lessons-from-a-real-world",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-14T01:59:31.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/jubilee_happy_a567008f769/from-claude-project-to-hybrid-ai-agent-lessons-from-a-real-world-content-workflow-562d"
  },
  "original_language": "en",
  "account": "The author initially used Claude to assist in preparing daily spiritual messages in English, but found that the workflow gradually evolved into a hybrid system. The weekly workflow begins with six Chinese daily spiritual messages, which must be translated, formatted, and published in both Chinese and English versions on the church website.\n\nThe process involves several steps, each requiring a different type of skill:\n\n1. **LLM skills**: The LLM (Claude) handles language and reasoning tasks, such as translating Scripture references, selecting appropriate terminology, structuring content semantically, and formatting elements like headings and subheadings. The LLM also produces output according to a specific interface contract, resulting in a predictable JSON structure.\n\n2. **Software skills**: Once the LLM generates the JSON output, a Python script takes over to insert the English content into an English PowerPoint template. This step requires deterministic transformations, where the LLM's output is simply placed into specific text boxes, and the software knows the mapping between days and corresponding slides. However, human review is still necessary due to the visual nature of PowerPoint formatting, such as adjusting text fit, paragraph breaks, and font adjustments.\n\n3. **Human skills**: After the English PowerPoint is reviewed, another automation step merges the English and Chinese PowerPoints into a single presentation containing twelve messages aligned by date. This task does not require an LLM, as the correspondence between Chinese and English slides is straightforward. The author then emphasizes the importance of human judgment in translation, theological review, visual quality control, and final publishing approval.\n\n4. **Final transformation**: The last step converts the final English PowerPoint into HTML for publishing on the church website. Initially, this task may seem like a job for an LLM, but it actually requires preserving specific formatting elements like paragraphs, titles, bold text, italic text, and ordering. While a program can extract the PowerPoint text and generate basic HTML, an LLM could assist with semantic interpretation, and a human can review the final result before publishing.\n\nThe author concludes that the AI system in this real-world content workflow is not a single model, but an orchestrated combination of different capabilities. By understanding the distinct skills required at each step, the workflow becomes more efficient and effective, with a clear separation between automation and human judgment.",
  "summary": "When I first started using Claude to help prepare our church's daily spiritual messages, I was not trying to build an AI agent. I simply wanted to reduce the repetitive work involved in preparing six Chinese messages for publication in English. But after several iterations, I realized that I was gradually designing something that looked much more like an agentic workflow: an LLM handled language…",
  "key_points": [
    "Claude assists in translating and formatting Chinese daily spiritual messages.",
    "Python script handles deterministic transformations for PowerPoint formatting.",
    "Human skills crucial for translation, theological review, and final publishing approval."
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}