user-persona-synthesizer

User Persona Synthesizer

This skill extracts recurring user profiles from real CustDev interview transcripts and generates structured, evidence-backed persona cards. It works from any free-form text or markdown files — no integrations, no special format required.

Input:

• Interview transcripts or notes: pasted text, markdown files, or plain text (any format accepted)
• Optional: focus attributes (e.g., "focus on goals and frustrations only") or persona count preference

Output:

• Markdown document with a summary table, per-persona cards (with verbatim quotes), and synthesis notes

---

Language Detection

Detect the user's language from their message:

• If Russian (or contains Cyrillic): respond in Russian
• If English (or other Latin-script language): respond in English
• If ambiguous: respond in the language of the trigger phrase used

---

Instructions

Step 1: Validate and Parse Input

1. Check that interview content is provided (pasted text or file reference)

   • If input is empty or whitespace only: Stop. Return: "No transcripts provided. Paste interview notes or reference files to begin persona synthesis."

2. Detect input format

   • If input resembles structured data (CSV column headers, spreadsheet rows): flag it — "Structured data detected — extracting patterns from columns. For richer personas, raw interview text works better." Proceed.
   • If input is clearly not interview content (product spec, article, meeting minutes with no respondent voices): Stop and report — "This doesn't look like interview transcripts. Persona synthesis requires user interview content — notes, quotes, or dialogue with respondents."
   • If user explicitly requests synthetic/hypothetical personas (not from real data): Stop and report — "This skill synthesizes personas from real interview data only. For synthetic personas, ask Claude directly without this skill."

3. Count the number of distinct interviews or respondents

   • Use explicit separators ("Interview 1:", "Respondent:", "---", numbered sections) or infer from context
   • Report count: "Found N interview(s). Proceeding with synthesis."
   • If only 1–2 interviews detected: flag low sample — "Low sample size (N interviews) — patterns may not be representative. Consider adding more interviews before sharing with stakeholders." Continue.