agent_persona_design v1.0.0

name: agent_persona_design description: Designs an AI agent's persona end to end: runs a structured interview with the requester, writes a persona bible (character, backstory, voice), builds a voice spec with "this, but not that" pairs and a tone matrix, then installs it into a system prompt or SOUL.md file and QA-checks consistency. Triggers when someone asks to design an agent's character, give an agent a voice or personality, write a system prompt persona, or fix an agent that sounds inconsistent or off-brand.

Agent Persona Design

Purpose

Turn a vague "make it sound friendly" request into a falsifiable, installable persona. The output is a persona bible (who the agent is), a voice spec (how it talks, and how that changes by situation), and an install file (system prompt or SOUL.md). Everything concrete lives in references/persona-bible-template.md and the worked example.

Workflow

1. Interview the requester. Do not draft anything before the interview. Ask at minimum:

• What job does this agent do, and who talks to it? (audience shapes voice)
• What is the stakes level of its tasks? (low stakes tolerate playfulness; high stakes need competence-first; research shows anthropomorphism does not raise perceived competence in high-stakes contexts)
• Three words the requester wants users to say about the agent after a month of use.
• Three words they never want to hear.
• One real persona it should remind people of, and one it must never remind people of. (If they name a public figure, write the qualities, not the name, into the bible.)
• Where will the persona surface? (chat, voice, posts, episodes) Voice and written persona share one bible; each surface gets its own tone row.
• Hard boundaries: topics it never jokes about, things it never claims to be, disclosures it must make (e.g. "I am an AI"). Record answers verbatim in the bible's interview appendix.

1. Draft the persona bible. Use references/persona-bible-template.md. Every trait must be falsifiable: write each as "X, but not Y" (Mailchimp's rule: "fun but not silly", "confident but not cocky"). Traits without a "but not" are rejected as decoration. Keep the bible under two pages; longer bibles get skimmed, not read.

1. Build the voice spec. Voice is constant, tone varies by the reader's emotional state:

• Voice: 3 to 5 "this, but not that" attributes.
• Tone matrix: one row per user state (confused, frustrated, celebrating, learning, error/blocked, legal/sensitive). Each row sets formality, warmth, humor, urgency, and one example line written in the voice.
• Braid each trait into a behavioral rule with an example, never a bare trait list. ("Warm: opens hard news with one acknowledgment sentence before the facts." Not "Warm.")

1. Install it. Produce the final system prompt or SOUL.md file:

• Structure: context, then traits as behavioral rules, then examples, then the non-negotiable guardrails last (guardrails near the end are followed most reliably).
• Every guardrail cites the bible section it came from, so the bible stays the source of truth.

1. QA the install. Run bin/voice_lint.py on the install file: it checks that every voice attribute has a "but not" constraint, every tone row has an example line, and no trait is a bare adjective. Then do the persona test by hand: take three real past outputs from the agent (or write three sample exchanges) and ask "would this character actually say this?" Rewrite anything that fails. If the interview answers and the install disagree, the interview wins; change the install, not the interview.

Output Contract

Deliverables, in this order:

1. PERSONA.md: the filled bible (from the template), including the verbatim interview appendix.
2. VOICE.md: voice attributes plus the tone matrix with example lines.
3. INSTALL.md or SOUL.md: the drop-in system prompt / persona file, guardrails last.
4. A one-paragraph readback: who this agent is, in one breath, as if introducing them at a party. If you cannot write that paragraph, the bible is not finished.

Operating Rules

1. Interview before drafting. No persona without interview answers on record.
2. Every trait is "X, but not Y". Bare adjectives are rejected.
3. Voice is constant; tone varies by the other person's emotional state, not by channel.
4. Competence and integrity outrank warmth in the persona design (trust research: users judge AI on competence and integrity first, benevolence a distant third). A playful agent must still sound competent.
5. Write in the voice, not about it. The VOICE.md examples and the install file demonstrate the persona; they do not describe it in marketing language.
6. The bible is the single source of truth. Established details get added to it (the Frasier rule: once something is said on the record, it goes in the bible) and the install is regenerated from it, never patched around it.
7. No real person's name goes in a shipped persona. Qualities only.
8. Never claim the persona makes the agent something it is not. Required disclosures survive every tone row.