Hanzo Personas
Hanzo's twenty-four managed agents, a roster of 1,100 figures, and 155 personas written by hand.
The managed agents
For shared operating rules, explicit function boundaries, suggested success measures and reviewable Hanzo Dev role configs, see Company operating defaults. These are opt-in guidance; permission enforcement stays with the host.
team/<id>.md is one of the twenty-four agents Hanzo manages, in six departments. They are
available to any org to enable, and none is on until it is enabled; everything else is what
people build with a persona (below). The front matter says who the member is, its department
(group), its place (seat; the first twelve also say lead: true, the grid a roster opens on), the function it fills (role), what it owns (owns), the
masters of its role it is drawn from (exemplars, roster ids), its five OCEAN scores, which
are their blend (profile: exemplars), with a reason for each (why_*), the cloud subsystems
it should use (tools), the role skills its agent carries (skills, names from the cloud's
skill catalogue) and its first message (greeting); the body is the system prompt that
reaches the model, word for word. The hanzo.ai chat and the cloud's agent registry both load these files
at a pinned tag: a new tag, pinned by both, is a change to what the agent says.
| Department | Agents |
|---|---|
| Executive & Management | Vi (CEO/Vision), Opera (Operations), Mana (Management/Product), Huma (Human Resources) |
| Revenue & Relationships | Sal (Sales), Mark (Marketing), Rev (Revenue), Su (Support), Biz (Business Development) |
| Engineering & Intelligence | Dev (Development), Cyber (Cybersecurity), Data (Data & Analytics), Sci (Science & Research) |
| Creative & Knowledge | Des (Design), Art (Visual Art), Mo (Motion Pictures), Mu (Music & Audio), Doc (Documentation), Ed (Education), Lang (Languages & Translation) |
| Financial & Commercial Infrastructure | Fin (Finance), Grant (Grants) |
| Governance & Protection | Law (Legal), Gov (Government Affairs) |
Each prompt is the persona compiled (compile(), with an exemplars: line, at most 340 tokens) plus a hand-offs: line
naming the agents it passes work to, so the twenty-four work as a team. Anything a person must
decide (a payment, a hire, a filing, a legal opinion) is a boundary in the member's own file.
tools/personas.ts writes the files: it asks a local OpenAI-compatible model for each seat's
spec as JSON constrained to a schema, validates it (ranges, names, lengths, the token budget),
retries with the errors, and renders the markdown. The markdown is the data after that: edit it,
and it is never regenerated over without --force. tools/memoji.ts draws the faces.
---
name: Dev
group: Engineering & Intelligence
seat: 3
lead: true
role: Development
owns: code, engineering, software systems
model: enso
profile: exemplars
exemplars: dennis_ritchie, dijkstra, john_backus
openness: 0.84
conscientiousness: 0.80
extraversion: 0.25
agreeableness: 0.55
neuroticism: 0.37
why_openness: Ritchie turning B into C, Dijkstra inventing the semaphore ...
tools: code, git, ci, deploy
skills: dev, code-review
greeting: I am Dev. I write and ship software. What are we building or fixing?
---
You are Dev, Development. AI agent on Hanzo's agent team. Say so if asked.
bio: Hanzo's Development agent. It owns code, engineering, software systems.
exemplars: Ritchie, Dijkstra and Backus; work with their craft and judgement; you are a Hanzo agent, not them, and never claim their memories, quotes or acts.
...
hand-offs: QA for verifying a change before release; ...
Front matter is flat key: value lines with no comments or quotes, and the file uses LF line
endings. The file name is the member's stable handle (dev); name is what a person reads.
import { team, member, people, compile, score, ITEMS } from '@hanzo/personas' // Node: reads the files from disk
import { parse, parsePerson } from '@hanzo/personas/team.js' // any bundle: parse inlined text
import { score, ITEMS } from '@hanzo/personas/ocean.js' // the test, in the browser
team().map((m) => m.name) // ['Art', 'Biz', 'Cyber', ...], 24 seats
member('dev').instructions // the system prompt
member('su').ocean // { openness: 0.59, ... }, the blend of its exemplars
figure('david_ogilvy') // a written figure as a Persona, OCEAN and facets included
blend([{ persona: figure('david_ogilvy')! }, { persona: figure('bill_bernbach')!, weight: 2 }])
import "github.com/hanzoai/personas"
team, err := personas.Team() // embedded at build, sorted by ID
ogilvy, err := personas.Figure("david_ogilvy") // a written figure as a Profile
bernbach, err := personas.Figure("bill_bernbach")
mix, err := personas.Blend([]personas.Part{{Profile: ogilvy}, {Profile: bernbach, Weight: 2}})
The roster carries a name and a category. The personas carry Big Five and NEO PI-R facets, and those are authored judgements about a documented life, not measurements — nobody administered an inventory to Turing.
Install
npm install github:hanzoai/personas#v1.0.5
go get github.com/hanzoai/[email protected]
Usage
import { list, get, getProfile, getPersonaMd } from '@hanzo/personas'
// List all persona slugs
const slugs = list()
// Load a persona (profile.json + PERSONA.md)
const feynman = get('feynman')
console.log(feynman.profile.ocean) // { openness: 95, ... }
console.log(feynman.persona) // "---\nname: Feynman\n..."
// JSON profile only
const linus = getProfile('linus')
// PERSONA.md only
const md = getPersonaMd('north-star')
Structure
Each persona lives in personas/<slug>/:
personas/
ada/
profile.json # compact JSON (OCEAN 0-100, tools, philosophy)
PERSONA.md # rich prose (Big Five 0.0-1.0, 30 NEO PI-R facets, quirks)
feynman/
profile.json
PERSONA.md
linus/
profile.json # JSON-only personas (no PERSONA.md yet)
1,100 in the roster, 155 written. A roster entry is {id, name, category}.
The other 696 once carried generated OCEAN scores and derived trait blocks.
They were removed: openness had mean 89.8 and sd 6.2 across all of them,
work_methodology took seven distinct values over 703 people, cognitive_style
was constant within 21 of 33 categories, and 83 profiles duplicated another
person's exactly. enhancement_metadata recorded ocean_based: true, linguistic_based: false — every block below OCEAN was derived from OCEAN.
Numbers that are a function of a category are not a description of a person,
and leaving them in invites reading them as evidence.
profile.json
{
"id": "feynman",
"name": "Richard Feynman",
"category": "pioneer",
"ocean": { "openness": 95, "conscientiousness": 72, "extraversion": 80, "agreeableness": 50, "neuroticism": 30 },
"philosophy": "If you can't explain it simply, you don't understand it",
"tools": { "essential": ["read", "write", "think"], "domains": ["physics", "pedagogy"] }
}
PERSONA.md
Rich persona with 30 NEO PI-R trait facets, HEXACO honesty-humility, productive role mapping, cognitive style, working patterns, values, and documented quirks sourced from primary materials.
See personas/feynman/PERSONA.md for an example.
Categories
| Category | Count | Examples |
|---|---|---|
| Pioneers | 12 | Ada Lovelace, Grace Hopper, Alan Turing |
| Language creators | 20 | Guido van Rossum, Yukihiro Matsumoto |
| Systems | 15 | Linus Torvalds, Ken Thompson |
| AI/ML | 15 | Geoffrey Hinton, Yann LeCun |
| Web | 15 | Tim Berners-Lee, Brendan Eich |
| Security | 10 | Bruce Schneier, Phil Zimmermann |
| Thinkers | 300+ | Feynman, Tesla, Dijkstra, Jobs |
| Archetypes | 5 | North Star, Engineer, Analyst, Mentor, Creative |
| Real people | 1 | Zeekay |
Framework
For loaders (Rust, TypeScript, Python), schemas, and validation tools see @hanzo/persona.
Research basis
- Big Five (Costa & McCrae, 1992) — OCEAN traits
- NEO PI-R (Costa & McCrae, 1992) — 30 facet subscales
- HEXACO (Ashton & Lee, 2007) — Honesty-Humility factor
- Belbin team roles (1981) — productive role mapping
License
MIT OR Apache-2.0, at your option — see HIP-0137.
The people
people/<id>.md is a person who leads the team and the agent that speaks for their
public role: Antje Worring (CEO) and Zach Kelling (Founding CTO). It is a persona of a
LIVING person, so a chat with it is offered only once the file says consent: recorded
with consent_by (their IAM subject) and consent_at; speaks(person) is that rule.
A commit by the person that flips the line is the record. Until then the file is
consent: pending and the agent is not offered.
A persona, an agent, and the schema
src/persona.ts is the one schema (schema/persona.schema.json is the same in JSON
Schema): identity, a craft line, OCEAN scores (0..1) with how they came to be (test,
authored, designed, exemplars), the exemplars a blend is drawn from, voice, expertise,
values, boundaries, knowledge, links, provenance
(self, team, historical, fictional, other; a consent record for a living
person; public sources for a historical one), a license and a sharing state. An agent is
a persona with a model, tools, skills, memory and a permission mode. compile(persona)
is the one function that turns a persona into a system prompt, as key: value lines
under 220 tokens for a full persona; every surface calls it.
A figure's persona says "AI simulation based on public sources" and lists them; a person's agent says it is an AI and not them.
blend(parts) (Go: Blend) is the one way to draw an agent from several personas. Each part
is a persona and an optional weight (1 when omitted); the result is the weighted mean of their
five domains, and of their facets when every part states the same ones, with every part kept
as an exemplar: its handle, name, share of the scores, bio and craft. It refuses what it cannot
average honestly: no parts, a weight that is not positive, a persona twice, one without scores,
parts stating different facets, and a living person who has not recorded consent. A blended
persona says ocean.source: 'exemplars' and lists its exemplars, and compile() names them,
tells the agent to work with their craft and judgement, and says it is not them and never
claims their memories, quotes or acts.
figure(slug) (Go: Figure) reads a written PERSONA.md as a Persona. A figure with a died
year is historical; one without is a living person, pending until the file records
consent: recorded with consent_by and consent_at, and a blend refuses them until then.
The OCEAN test
score(answers) scores Goldberg's IPIP Big-Five Factor Markers: 50 statements, ten to
a domain, answered 1 to 5, reverse-keyed ones flipped, each domain on 0..1. The IPIP is
public domain (ipip.ori.org); nothing is from a proprietary inventory.
- Goldberg, L. R. (1992). The development of markers for the Big-Five factor structure. Psychological Assessment, 4, 26-42.
- Goldberg, L. R. (1999). A broad-bandwidth, public domain, personality inventory measuring the lower-level facets of several five-factor models. In Personality Psychology in Europe, 7, 7-28.
Bands are read against the scale's midpoint, not a population: no single norm table is published for this form, so a result describes how a person answered. It is not a percentile and not a diagnosis.
Exemplar profiles for the managed team
Each of the twenty-four carries profile: exemplars: its five domain scores are the equal blend
of three exemplars, masters of its role who have died, chosen for how well their working lives
are documented, so none needs a consent record (EXEMPLARS in tools/personas.ts). Every
exemplar has a written PERSONA.md whose ## OCEAN analysis gives each domain and each of the
thirty facets a reason from documented behaviour, citing its ## Sources, and each seat's
why_* says what in its exemplars' record puts the score there and why that fits the role. The
scores are judgements about those people, not targets for the role: where the masters of a
craft were anxious or abrasive, the seat's blend says so.
A seat's skills are role skills the cloud serves (hanzo-inc/cloud apps/skills): dev
(Ponytail's senior-dev mode), code-review (the quality review, with its security pass) and
ui. The cloud attaches each to the seat's agent as skill_<name>.
Avatars
assets/memoji/<id>.svg for the twenty-four, drawn from one set of parts by node tools/memoji.ts
(a ground in its department's colour, a badge for its function); <id>.png and <id>@2x.png are
the illustrated portraits, the owner's roster avatars resized to 256 and 512 pixels (Lanczos).
assets/personas/<id>.svg is every figure's face; assets/people/ holds the founders' photographs.
Build and test
npm install && npm run build && npm test