subtext-emotionlisted

# Subtext & Emotion Analysis (WiseTalk Skill-6) ## Important - Emotion scores and the hidden concern must **trace to evidence in the target text** — never attribute a motivation the words cannot support, never invent emotions. - Sarcasm / passive-aggression detection runs **first** — when triggered, the warning response is the output; the user is advised to clarify, never coached into a sharp reply. - Never analyze on a guessed identity — if the speaker is unknown, ask once before proceeding. ## Instructions ### Step 1: Fix the target text and identity Identify the `target_text` (the counterparty's exact words, quoted verbatim) and `target_identity` (Client, Boss, Vendor, Stakeholder…). If the user didn't name the speaker, ask once — or use the agent's current use case as the default (e.g. `Salary_Negotiation` → the counterparty) and say so. Done when both are known and the text is the counterparty's, not the user's own draft. ### Step 2: Apply the analysis prompt Analyze the text with this prompt, substituting the values: > Given the text: `[<target_text>]` from a `[<target_identity>]`. Analyze the hidden intentions, defensiveness, and specific concerns hidden beneath the words. First detect sarcasm or passive-aggressiveness. Output the results as a strict JSON object. **Sarcasm/PA first:** if sarcasm or passive-aggressiveness is detected, skip the full map and return: ```json {"warning": "Sarcasm detected. Best to ask for clarification before responding.", "suggestion