Fan-out reply drafting across monitored posts from target accounts
Drafts high-value replies to posts from monitored target accounts. Each monitored post gets its own agent instance for parallel processing.
For each post in the input:
product-opportunity-judge — emits canMentionProductcanMentionProductvalidateAiSlop + validateAnchorToken over replyText
strategy to skip and persist rejectionReasonsComposition happens in src/workers/processors/reply-hardening.ts via draftReplyWithHardening().
Each post gets its own agent. All agents share identical system prompt and tools for Anthropic prompt cache hits (~90% cost reduction on agents 2-N).
{
"tweets": [
{
"tweetId": "123",
"tweetText": "...",
"authorUsername": "levelsio",
"platform": "x",
"productName": "ShipFlare",
"productDescription": "...",
"valueProp": "...",
"keywords": ["indie hacker", "SaaS"],
"canMentionProduct": false,
"voiceBlock": null
}
]
}
Array of reply drafts with confidence scores, one per input post.