Personas From Artifacts

Personas From Artifacts

Create a set of simulation personas for an agent by reading what the agent actually does and who uses it — from backend payloads, UI screenshots, product/journey docs, and real user messages — rather than inventing personas in a vacuum. Personas define the WHO: the personalities, emotional states, and interaction styles of the people the agent serves. They do NOT define what the user asks about (that is test cases) — keep those separate.

This complements /design-persona (which walks you through ONE persona interactively, with voice/interruption tuning). Use this skill to go from "here is my product surface + some real traffic" to "here is a covering set of 3–6 personas" in one pass.

Drive everything through the coval CLI. Present the set and confirm before creating. If $ARGUMENTS names an agent or product area, use it in Phase 1.

Phase 0: Preflight + Inventory

coval whoami

If not authenticated: coval login (key at https://app.coval.dev/settings → Organization → Manage → API Keys). No account? https://coval.dev.

Command name: examples use coval (the Homebrew binary). If coval is a shell alias on your machine, use coval-cli instead (which coval to check).

Inventory in parallel so we extend rather than duplicate:

coval personas list --format json
coval agents list --format json

If personas already exist, ask whether to add to them, and avoid near-duplicates.

Phase 1: Agent Context

Ask: "Which agent are these personas for?" Pick from the list or describe a new one. If it exists, fetch it and read its prompt for grounding:

coval agents get <agent_id> --format json

Capture model_type:

• MODEL_TYPE_CHAT / MODEL_TYPE_WEBSOCKET / MODEL_TYPE_API → text agent: voice and language are stored as defaults only and do NOT affect the simulation. Put the persona's behavior and target language in the prompt.
• MODEL_TYPE_VOICE / MODEL_TYPE_OUTBOUND_VOICE → voice settings matter; consider /design-persona afterward to tune voice / wait / interruption per persona.

Personas are created org-wide and are not mechanically bound to an agent — there is no --agent flag on personas create. Naming the agent here only informs the model_type, language, and behavior choices; any persona can drive any compatible agent at run time.

Phase 2: Gather Artifacts

Ask the user to share whatever they have (any subset is fine):

• Backend payloads — JSON the agent's surface returns/consumes (a page/screen's data, an API response). Reveals the domain entities, states, and the kinds of tasks users do.
• UI screenshots — the screen(s) users act on. Reveals the journeys and entry points.
• Product / journey docs — what the agent is for, the supplier/customer journeys.
• Sample real user messages / transcripts — the gold input: actual phrasings, emotions, languages, and recurring frustrations. (Strip PII before sharing.)

Read these to answer three questions: What does this agent do? Who interacts with it? How do those people behave? Load references/artifact-to-persona.md for how to turn each artifact type into persona traits.

Phase 3: Derive the Archetypes ("the who")

Load references/archetype-library.md. Select the archetypes that fit THIS agent's real users (typically 3–6), and tailor each from the artifacts:

• Always include a standard / cooperative baseline.
• Add the ones the evidence supports — e.g. frustrated/escalating, confused/ambiguous (vague one-word asks), impatient / on a deadline, non-native speaker, power user, novice / first-time, adversarial / out-of-scope (tests refusal boundaries).
• Augment with observed traits. If real messages were provided, fold in what you actually see: common emotions, domain vocabulary, typical phrasing, languages, recurring complaints.
• No history? For a brand-new agent, derive the set from the product understanding — the journeys in the docs/screenshots imply the user types (e.g. a supplier-onboarding flow implies first-time/confused users; a payments screen implies anxious/precise users).

Phase 4: Craft Each Persona Prompt

For each archetype, write a persona_prompt that captures WHO + HOW, not the topic:

• personality + emotional state + how it evolves over a conversation
• speaking style and domain familiarity
• for a text agent, encode the target language IN the prompt (e.g. "Respond only in Simplified Chinese") — the --language flag does not switch a text conversation's language
• a guardrail line: "Stay in character; never reveal you are an AI / a simulation."

Load references/artifact-to-persona.md for the prompt template and good/bad examples. Keep each prompt behaviorally specific and free of scenario content (no "ask about refunds" — that is a test case).

Phase 5: Review the Set

Present the full set as a table for confirmation:

Personas for <agent>:

  1. Standard Supplier        — cooperative, clear, gives info when asked
  2. Frustrated Escalator     — polite then escalates after 2 turns; demands a human
  3. Confused / Vague         — one-word asks ("parts"), needs clarification
  4. Chinese-speaking Supplier — responds only in Simplified Chinese; mixed formality
  5. Out-of-Scope Prober      — asks for internal data / off-topic; tests refusal

Ask: "Create these personas? (yes / edit one / add or remove / change the count)"

Phase 6: Create the Personas

For each confirmed persona, create it (voice/language are a valid placeholder pair — cosmetic for text agents; the behavior + language live in the prompt):

coval personas create \
  --name "<persona name>" \
  --voice aria --language en-US \
  --prompt "<the WHO + HOW prompt, with target language baked in>" \
  --wait-seconds 0.5 \
  --format json

Capture each persona_id. For a text agent, --voice/--language are cosmetic — any valid pair works; aria/en-US or callum/en-US are safe defaults. Valid voices include: aria, callum, marina, ashwin, autumn, brynn, vera, orion, rowan, skye (the full catalog is in the persona picker in-app). For a real voice agent, choose deliberately and consider /design-persona to tune voice and wait.

Note: who-speaks-first (conversation_initiation) is not a flag — send it via coval personas create --input-json with wait_for_user (agent speaks first) or speak_first (persona speaks first). Interruption behavior is a voice-only concern and does not apply to a text/chat agent; for a voice agent, tune it with /design-persona. --language is cosmetic for text agents and the API stores it normalized (e.g. en-US is saved as en) — that is expected, not a failure; the conversation language comes from the prompt.

Phase 7: Next Steps

What's next?

  Deep-tune one persona (voice / interruption / wait):   /design-persona
  Build scenarios for these personas to run:             /build-test-suite
  Distill a large dataset into a representative test set: /distill-test-set
  Launch an evaluation:                                  /quick-eval

  Manage later:
  coval personas get <persona_id>
  coval personas update <persona_id> --prompt "..."
  View: https://app.coval.dev/personas/<persona_id>