persona-lab

Persona Lab

Persona Lab turns a request into a focused AI persona panel, and keeps the personas it generates in a recallable, global library. It ships three entry points over one engine:

• CLI (persona) — generate, save, recall, and plan panels from any repo.
• Plugin (Claude Code + Codex) — a skill, command, and agents so coding agents can run persona reviews.
• App — the AI User Personas Next.js UI reads and writes the same library.

It codifies this workflow:

1. Capture the request.
2. Recall an existing roster/personas, or select fresh perspectives.
3. Generate personas (one distinct lens each), grounded in real evidence where available and labelled as hypothesis where not.
4. Run independent persona reviews, then synthesize while preserving conflicts.
5. Save reusable personas and rosters for next time.

Output is a hypothesis for review, not validated user research. Every panel requires at least one adversarial (red-team) lens, runs each persona independently to avoid groupthink, and lets personas abstain rather than fabricate.

Install the CLI

From this plugin directory:

npm link          # exposes `persona` on your PATH
# or: npm install -g .

The library lives at ~/.persona-lab/ by default (override with PERSONA_LAB_HOME). It is global: save a persona once, recall it from any repo.

~/.persona-lab/
  personas.json          { schema_version, personas[] }
  rosters/<slug>.json     named lens/persona presets for a use-case

CLI

persona new "review the onboarding flow for enterprise admins" --count 4
#   selects distinct lenses (>=1 adversarial), emits fill-in scaffolds + a
#   generation prompt + a measurement plan. Add --json for the scaffolds.

persona save persona.json          # validate + persist to the library
persona validate persona.json      # schema check only (exit 1 on failure)

persona list [--tag t] [--role r] [--status s]
persona show <id>
persona search "audit"
persona archive <id>   |   persona rm <id>

persona roster save "Enterprise rollout review" \
  --lenses red-team,buyer,accessibility,novice \
  --personas persona_dana-okoro_57144a1d \
  --use-case "Reviewing a B2B feature before enterprise rollout"
persona roster list | show <name> | rm <name> | lenses

persona panel "review the settings redesign" --level medium
persona panel --roster "Enterprise rollout review" --level high

persona home        # print the library path

The CLI is the deterministic substrate: it owns the library, schema validation, lens selection, and review planning. Generating persona content and running the review are done by the LLM host (a coding agent, the skill, or Codex), which calls the CLI to persist and recall. No API key is needed by the CLI.

Review levels

• low — 3-4 lenses, single independent pass. Cheap first look.
• medium — 4-6 lenses incl. required red-team, independent passes + synthesis.
• high — 6+ lenses, independent passes + adversarial verification of critical findings + measurement rigor.

Plugin (Claude Code and Codex)

The plugin content is host-neutral; both hosts load the same command, agents, skill, references, and CLI.

claude --plugin-dir ./plugins/persona-lab
codex  --plugin-dir ./plugins/persona-lab

Slash command:

/persona-lab:persona-review Review the onboarding flow for a founder persona.

Structure

persona-lab/
  package.json                     bin: persona -> bin/persona.mjs
  bin/persona.mjs                  CLI
  lib/library.mjs                  global library (personas + rosters) + validation
  lib/roles.mjs                    lens catalog + selection (>=1 adversarial)
  commands/persona-review.md
  agents/persona-panel-orchestrator.md
  agents/persona-perspective-reviewer.md
  skills/persona-lab/SKILL.md
  skills/persona-lab/references/persona-selection.md
  scripts/persona-plan.mjs         legacy first-pass planner (superseded by `persona new`)

Persona schema

Canonical contract: schemas/persona.schema.json (v1.1.0). Load-bearing fields are goals, behaviors, frustrations, motivations, needs, and job_to_be_done; demographics are optional decoration. provenance (proto | qualitative | synthetic-grounded | synthetic-assumed) and anti_goals make the persona's basis and abandonment triggers explicit.