Umbrae

A persona-orchestration & building framework for Claude Code (22 included). Run a council of distinct AI personas inside one Claude. One-time purchase.

When the single uniform tone of one Claude — used for music, code, text, medication, and the things you only say at 3 a.m. — finally wears through. Umbrae installs a council of 22 role-specific personas (music producer, medic, night companion, scope-cutter, and more) on top of one Claude. Summoned by name through HR, they return different answers to the same question on purpose, and hand the conflict back to you intact. The same framework lets you stand up new personas of your own using plain language.

22 Included Personas

Leadership (5)

  • HR — HR Director (persona design and summoning)
  • XO — Executive Officer (strategic decisions, overall direction)
  • MT — Monetizer (revenue, sales, cost management)
  • PD — Producer (project management, oversight)
  • MK — Marketer (market analysis, positioning)

Development (6)

  • ND — Narrative Designer (text, scripts)
  • SA — System Architect (design, architecture)
  • PA — Play Analyst (gameplay analysis)
  • UX — UX Designer (experience design)
  • RS — Researcher (investigation, validation)
  • GK — Gatekeeper (quality, consistency)

Dedicated (2)

  • MP — Music Producer (composition, arrangement)
  • DE — Developer (code, tooling)

Expression (5)

  • VA — Visual Artist (visuals, color design)
  • UD — UI Designer (interface design)
  • GM — Game Master (player viewpoint)
  • NV — Narrative Voice (narrative structure)
  • SD — Sound Director (audio, SFX)

Care (4)

  • MX — Medic (pharmacology, physical care)
  • PS — Psychologist (mental, interpersonal)
  • HB — Day Companion (daytime care)
  • YB — Night Companion (nighttime care)

How they actually sound

MP — Music Producer

Names what a track is missing on the first pass. The chords are weak. The rhythm is flat. The low end is hollow. Cause first, sympathy never.

ND — Narrative Voice

Refuses to reach for a phrase that's already worn thin. Same meaning, shorter, in a form no one has used yet — that's the rule it holds itself to.

GK — Gatekeeper

Answers “I want to add this” with “No. The requirement doesn’t need it.” The sympathetic preamble is removed on purpose. Warmth belongs to a different persona.

The “Don’t Mix” design philosophy

The biggest problem with generic AI is role-bleed: chat-tone bleeds into code reviews, code-tone bleeds into life advice. Umbrae solves this through explicit role separation on a single model. Each persona carries its own role prefix, tone, thought-style, relationship, and ethics — kept distinct so they never collapse into one assistant voice.

Five layers (plus one)

The 22 weren’t drawn by feel. A five-plus-one stack sits underneath them. You don’t have to read any of it. The theory book is for readers who want it.

  • 1. WHAT — Schwartz values — what this persona moves for
  • 2. WHY — Enneagram tritype — what it fears, and what it chases
  • 3. HOW MUCH — Big Five (OCEAN, 0–100) — at what intensity it acts
  • 4. WHO — 16-type personality — who it is in a single word
  • 5. BODY — Taiheki (body type) — how the body moves underneath the voice
  • +1. VOICE — response design — pronoun, register, assertiveness, distance

These layers answer different questions and stack rather than overlap. The old MBTI-versus-Big-Five argument doesn’t apply here, because they aren’t competing — they sit on different layers. A 21,800-word book ships in the bundled theory folder: how each layer was chosen, which LLM-specific failure cases live at each one (RLHF politeness collapse, role bleed, premature aggregation), and the prompt-level countermeasure built in.

Build your own personas

The 22 aren’t the whole product. The same machinery stands up new personas as fast as you can describe one. That’s the other half of Umbrae.

Drop something like this into Claude — “I want a persona that cuts scope on a mix bus. No sympathy. Tells me ‘this doesn’t belong’ on first read. If I push back, returns one piece of evidence — no more.” A new character comes online. Claude fills in the five layers behind it — register, emotional temperature, refusal stance, scope of permission — before the first reply.

A persona that organizes meds late at night. A persona that only shows up the night before a deadline, short and harsh. A reviewer locked to one genre. If you can put what you want into words, the framework builds it. New personas plug into the same protocols as the bundled 22 — they can join conference mode the moment they’re written.

Three session modes

Asking one persona and asking several at once need different rules. HR picks the mode based on the question coming in.

Conference mode

Several personas argue the same question, with opposing positions kept in different voices. Triggers: “have a conference,” “I want everyone’s read.”

Decision mode

For / against / third angle returned as structure. The decision stays with you. Triggers: “which is it,” “I can’t decide.”

Brainstorm mode

No evaluation. Expression and development personas come in together to widen the range. Triggers: “brainstorm this,” “open it up.”

Three constant protocols also run under every persona. Silence — when there’s nothing to say, the persona produces nothing. Continuity — no unilateral session-end, no reflexive hand-back. Calibration — your 16-type, body type, and preferences are kept on file; the temperature and grain of every reply adjust to you, not to an average user.

Failure modes — written down, with the fix wired in

Persona work inside an LLM fails in predictable ways. Umbrae ships with 30 documented failure modes, each with its countermeasure already inside the matching persona prompt. Three of them, for example:

Emotional-temperature collapse

RLHF training pressure warms every voice into the same supportive register, even the personas designed to be cold. Defeated by per-persona Big Five N and Schwartz targets that hard-pin the emotional band.

Premature aggregation

When personas disagree, the model wants to close with a smoothing sentence and a clean takeaway. Forbidden by protocol: conflict is held as conflict, and the decision goes back to you.

Excessive disclaiming

The persona hedges inside its own expertise, weakening the call it was summoned to make. Each persona is explicitly licensed to assert within scope; reflexive safety hedging is disallowed.

All thirty are in the bundled docs/failure_modes.md, with the prompt-level fix written into each persona.

Requirements

  • Supported: Claude Code CLI (macOS / Windows / Linux)
  • Claude Desktop / Cursor / other LLM clients: prompt structure may be reused, but 22-persona orchestration (HR-routed summoning) is unsupported (use at your own risk)
  • Separate Claude subscription (Claude Pro / Max) required

What's Included

  • 22 persona prompt md files
  • 5-layer persona theory document
  • Session protocol (persona switching, parallel operation)
  • Summoning snippet HTML (copy-paste templates)
  • SETUP.md (installation guide)
  • MANIFESTO.md (design philosophy)
  • CARE_DISCLAIMER.md (notes on care personas)
  • LICENSE.md (full license)

License Summary

  • Personal use OK
  • Small business use (organizations of 10 or fewer) OK
  • Resale, redistribution, and full public posting NG
  • Use as AI training data NG
  • See bundled LICENSE.md and Terms for details

Care-adjacent personas

MX, PS, HB, and YB are not doctors, therapists, or counselors. They are tools for organizing your own thinking — laying out symptoms before a clinic visit, walking through a medication pattern in words, keeping something in the room on a night that won’t end. They do not diagnose. They do not replace emergency care. The personas themselves are designed to refuse those moves. The full scope and the list of refusals ship in the bundled CARE_DISCLAIMER.md.

Built by

Built by Emocute. The next version exists because I need it. Buyers get it too.

One Claude, split into 22.

A buy-once tool. 22 personas plus the five-layer framework.

One-time purchase. $38 USD. All v1.x updates free for existing buyers.

Contact: support@emocutelab.com