---
name: derive-personas
description: Derive evidence-based personas from patterns that recur across real interviews and survey results — bottom-up, every trait traceable to evidence
argument-hint: "[interview/survey files to draw from; defaults to all real-source material in product/]"
disable-model-invocation: true
---

# /derive-personas

Derive personas bottom-up from real research evidence. The opposite move from `/generate-personas`: that one invents archetypes top-down as hypotheses; this one clusters patterns that actually showed up in the data.

Input: $ARGUMENTS

## Workflow

1. **Gather the evidence.** From the named files, or by default: transcripts in `product/interviews/` marked `source: real`, survey analyses in `product/insights/` marked `source: survey`, plus any raw material the user points to. If the corpus is thin (fewer than ~4–5 real interviews and no survey data), say so — patterns from 1–2 conversations are anecdotes — and offer to proceed anyway with the personas explicitly marked low-confidence.

2. **Cluster by behavior, not demographics.** Read the corpus and group interviewees/respondents by recurring patterns in goals, pains, workflows, and context — two people with different job titles who share the same struggle belong together; same title, different struggle, apart. Present the candidate clusters with their supporting evidence and rough coverage (how many sources support each) before writing any persona.

3. **Draft one persona per confirmed cluster,** using the structure from the `synthetic-personas` skill, with two differences: every major trait must be traceable to the evidence (attach 1–3 supporting quotes per section — never fill gaps with invention; a section without evidence stays marked "no evidence yet"), and the header carries `source: derived` plus the list of source files with coverage. The `type:` field is assigned from the evidence, never invented: a cluster that suffers the problem and would use the product is `primary`; one that uses it occasionally or for something else, `secondary`; one that does not use the product but decides, pays, or approves it, `tertiary`; one that shows up in the data and falls outside the declared scope, `negative`. When the evidence does not settle it, the type stays marked "no evidence yet" like any other section.

4. **Reconcile with existing personas** in `product/personas/`. For each existing synthetic persona: validated by a cluster (note it and suggest merging the evidence in), contradicted (say how, and suggest revising or retiring it), simply not observed in the data yet (leave it, noted), or **typed wrong** — the type the synthetic persona assumed does not hold in the data (a supposed `primary` that never uses the product but signs off on it is a `tertiary`): propose the type change with the evidence behind it; if the user accepts, update the persona's `type:` line and record the correction in `product/corrections.md` (dated entry: artifact, what the AI assumed, what the human decided, why). An existing persona with no `type:` line gets one proposed from the evidence the same way. The user decides; never delete anything unasked.

5. **Present the set, iterate, and save** each persona to `product/personas/{name-slug}.md`. Derived and synthetic personas live in the same directory, distinguished by the `source:` header.

6. **Close in one line:** the whole loop works with derived personas too — interview them (`/interview-persona`), run critique panels (`/critique-spec`) — now with the personas standing on evidence instead of hypothesis.

## Language

Conversation and the saved personas in the language of the conversation (matching the existing persona files).