{
  "id": 20400565,
  "title": "Stop asking one prompt to do three jobs",
  "url": "https://urgent.news/2026/11/17/stop-asking-one-prompt-to-do-three-jobs",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-17T07:51:41.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mgazar_/stop-asking-one-prompt-to-do-three-jobs-lk0"
  },
  "original_language": "en",
  "account": "When tackling complex AI workflows, it's crucial to avoid the common mistake of attempting to accomplish multiple tasks with a single prompt. This approach often leads to failure in edge cases, as different goals can inadvertently compete with each other, causing issues that are difficult to identify. The root cause of such problems lies in the design, not the model itself.\n\nTo address this issue, it's recommended to divide the work among multiple agents, each assigned a single specific task. For example, a compliance review that used to be handled by one prompt can be broken down into three distinct agents. The first agent, an extraction agent, is responsible for pulling out relevant facts from the source material and organizing them into fixed fields. This agent should never state facts that are not present in the source, use null placeholders for missing fields, make judgments about importance, rewrite or summarize the source, or add any fields outside the specified schema.\n\nThe second agent, an assessment agent, evaluates whether the extracted facts meet a certain threshold, such as materiality. This agent should only work with the extracted fields and have no exposure to the raw document, ensuring that its tone and language do not influence the decision-making process. The third and final agent, a drafting agent, generates the response based on the assessment made by the assessment agent.\n\nWhen designing these agents, it's essential to use anti-goals rather than trying to define personas. An anti-goal is a list of things the agent must never do. For instance, the extraction agent should never state a fact not found in the source, make judgments about importance, rewrite or summarize the source, or include fields not defined in the schema. By explicitly listing these anti-goals, the agent's behavior can be rigorously tested and validated.\n\nIt's important to avoid passing along the entire conversation history between agents. Doing so not only consumes tokens but can also reintroduce contamination, defeating the purpose of splitting the work. Instead, each agent should receive a fixed JSON payload containing the relevant information and return an output in the same format. At every handoff, the payload should be validated to ensure that all required keys are present and have the correct data types. If any validation fails, the process should stop to prevent the next agent from making assumptions based on incomplete or incorrect data.\n\nBefore writing any prompts for the agents, it's recommended to create an Agent Specification Matrix. This matrix should contain one row per agent, detailing its role, the scope of its operation, the anti-goals it must adhere to, the input it accepts, and the output it guarantees. By filling out this matrix first, potential design mistakes can be identified and corrected before moving on to writing the actual prompts.\n\nIn summary, when designing AI workflows that involve multiple tasks, it's best to split the work among multiple agents, each assigned a single transformation. By defining clear anti-goals for each agent, validating payloads at every handoff, and creating an Agent Specification Matrix beforehand, you can ensure that the system functions as intended and avoid the pitfalls of attempting to accomplish too much with a single prompt.",
  "summary": "Adapted from the first lesson of Designing Multi-Agent Systems on Learnvoro. Originally published on Medium . Most AI workflows start as one long prompt: read this document, decide what matters, write the response. It works on the easy cases. Then the edge cases arrive, and it fails in ways that are hard to spot: a summary that quietly includes a penalty the source never mentioned, or a judgment…",
  "key_points": [],
  "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."
}