Schema-validated LLM output with claude --json-schema

Pipe Reddit/HN/anything into claude with a strict JSON Schema — server-side validates the output shape. Downstream jq can trust every field.

$ curl -s -A "oneliner101/1.0" \
     "https://www.reddit.com/r/LocalLLaMA/top.json?limit=20&t=day" \
  | jq -r '.data.children[].data | "- [\(.score)] \(.title)"' \
  | claude -p \
      --output-format json \
      --json-schema '{
        "type":"object",
        "required":["theme","sentiment","top_post"],
        "properties":{
          "theme":{"type":"string"},
          "sentiment":{"type":"string","enum":["bullish","bearish","mixed"]},
          "top_post":{"type":"string"}
        }
      }' \
      "Summarize these posts: theme (one phrase), sentiment, top_post." \
  | jq -r '.result'

1. [01] curlcurl … reddit.com/r/LocalLLaMA/top.json …
   Reddit JSON, custom User-Agent required.
2. [02] jqjq -r '… | "- [\(.score)] \(.title)"'
   Tight prompt input — one '[score] title' line per post.
3. [03] claudeclaude -p --output-format json
   Returns a JSON envelope with .result, .session_id, .duration_ms, .usage. Pair with --json-schema or it's just stringified text inside.
4. [04] claude--json-schema '{ "type":"object", … }'
   Server-side validation. Claude's response is constrained to match — invalid shapes fail at the API, not in your shell. Downstream jq can trust every field exists.
5. [05] jqjq -r '.result'
   Extract the validated payload from the envelope. .result is now guaranteed to match your schema.

{
  "theme": "agentic-RAG with hybrid retrieval",
  "sentiment": "bullish",
  "top_post": "Llama 4 70B released, here are the first benchmarks"
}

• Both --output-format json AND --json-schema are required together. Without --output-format, the schema is ignored.
• Schema enforcement is strict — extra fields cause an error. Use `additionalProperties: false` defensively.
• Pipe straight into a database INSERT, a webhook, or a typed config — no defensive parsing.