{
  "$schema": "https://promptfoo.tech/schemas/workflow-template.json",
  "name": "Claude Research Brief: From Question to Cited Draft",
  "slug": "claude-research-brief-workflow",
  "platform": "Claude",
  "category": "claude",
  "difficulty": "Intermediate",
  "tags": [
    "claude",
    "research",
    "writing",
    "long-context"
  ],
  "summary": "A three-stage Claude workflow that turns a research question into a structured brief with sources, key claims, and a first draft you can edit.",
  "overview": "Long-context models make it possible to research and write in the same session without losing citations. This workflow uses Claude's 200k context to hold source material, extract claims, and produce a first draft that references the sources by number.",
  "variables": {
    "ANTHROPIC_API_KEY": "sk-ant-..."
  },
  "steps": [
    {
      "order": 1,
      "name": "Collect and label sources",
      "description": "Save each source as plain text or clean markdown. Number them S1..Sn. Strip navigation and ads."
    },
    {
      "order": 2,
      "name": "Paste with the framing prompt",
      "description": "Use the provided system prompt that instructs Claude to only cite sources present in the message and to refuse otherwise."
    },
    {
      "order": 3,
      "name": "Request the claims table",
      "description": "Ask for a markdown table with claim, supporting source(s), and confidence. Reject any row without a source."
    },
    {
      "order": 4,
      "name": "Generate the draft",
      "description": "In a follow-up turn, ask for the 800-word brief. Insist on inline [S3] style citations."
    },
    {
      "order": 5,
      "name": "Human edit pass",
      "description": "Read every citation. Anything not directly supported gets cut, not softened."
    }
  ],
  "howItWorks": [
    "Stage 1: paste 5–15 source articles into a single message with numbered delimiters.",
    "Stage 2: ask Claude to output a claims table (claim, source #, confidence).",
    "Stage 3: ask Claude to write an 800-word brief that only uses claims marked high-confidence and cites source numbers inline."
  ],
  "benefits": [
    "Cuts research-to-first-draft from a full afternoon to under an hour.",
    "Every claim in the draft is traceable to a source you provided.",
    "Reusable across topics by changing only the sources."
  ],
  "useCases": [
    "Analysts writing weekly market notes.",
    "Content teams producing pillar articles from vetted sources.",
    "Consultants preparing briefing docs for client meetings."
  ],
  "example": null,
  "faqs": [
    {
      "q": "Why not use web search?",
      "a": "Web-search agents drift and cite pages they didn't actually read. Providing sources up front produces more auditable output."
    },
    {
      "q": "How many sources can I include?",
      "a": "Practically 15–25 well-trimmed articles fit comfortably. Beyond that, chunk into topic clusters."
    },
    {
      "q": "Does this work with GPT-4?",
      "a": "Yes, but you'll want to reduce sources to fit its context window and be stricter about citation refusal in the prompt."
    }
  ],
  "source": "https://promptfoo.tech/workflows/claude-research-brief-workflow",
  "updated": "2025-06-02",
  "license": "CC-BY-4.0"
}