# Clone — Design & Strategy Pass (Fable) --- ## 1. Clone vs. Persona **A persona is designed; a clone is derived.** A persona answers "what should this role say?" — it's authored toward a goal (a brand voice, a helpful teacher, a spokesperson) and its obligation is to the goal. A clone answers "what would this person say?" — it's bound to evidence: their recordings, writings, cadence, and history — and its obligation is *fidelity to a real human*. That distinction is not cosmetic; it's the entire responsibility model. A persona that improvises is doing its job. A clone that improvises is *lying about a person*. So Clone the platform can host both, but the word "clone" is the right flag to fly, because it names the harder, more valuable, more dangerous thing: presence with provenance. Every design decision below flows from that — clones show their sources, disclose their gaps, and never claim to be more than the evidence supports. Use that paragraph publicly. It also pre-answers the critics: "How is this not deepfakes?" — *Deepfakes hide provenance. Clones are made of it.* --- ## 2. The Human-Interface Thesis (the soul) **Why humans need human interfaces:** A face and a voice are not a skin over information — they're a second channel. Tone, hesitation, warmth, a raised eyebrow at your question: these carry trust, permission, and emotional pacing that text cannot. For roughly half the population, a blank prompt box is not an invitation, it's an exam. The prompt box asks *you* to know what to say. A face asks nothing — you already know how to talk to a face. You've been doing it since before you could read. **Where:** at moments of vulnerability and first contact. Grief (a grandchild missing grandpa). Learning (a student afraid to look stupid). High stakes (a customer who wants to feel *addressed*, not processed). Onboarding to anything new — including AI itself. **The bridge claim, stated plainly:** Clone is not an AI product with an avatar bolted on. It is the *first AI surface that requires zero AI literacy*. No prompting, no settings, no jargon. You land, someone is there, you talk. That's the whole interface. For the hundreds of millions of people who will never open ChatGPT, Clone is the on-ramp — and for the people they love, it's the destination. The product's job is to make the AI *disappear behind the human*, while the RAI layer makes sure the human-ness never becomes a deception. **The one-line soul:** *You don't use Clone. You meet it.* Every screen should be gradeable against that sentence. If a screen looks like software, it's failing. --- ## 3. Use-Case Pontification **Grandpa's stories (emotional flagship).** *Scene:* Sunday evening. A nine-year-old on the couch says "tell me about the ship again." Grandpa — cloned from forty minutes of camcorder tape and two answering-machine messages — leans in, chuckles the way he actually chuckled, and starts the story of coming over from Palermo. The kid interrupts with a question the tape never covered: "Were you scared?" And Grandpa says, "You know, I told a story once about the storm on the third night — want to hear that one? It might answer you." *For:* families, especially those with aging parents *now*. *Moment of magic:* the interruption. A video can't be interrupted. Presence can. *RAI guardrail:* the **evidence boundary**. When asked something the source material doesn't cover, the clone never fabricates a memory — it redirects to what it truly holds ("I told a story once about…") or gently owns the gap. And it never claims to *be* him: it holds him. This distinction must be designed into voice and script, not buried in a disclaimer. **Company spokesperson avatar.** *Scene:* A prospect lands on a mid-size company's site at 11pm. Instead of a chatbot widget, the company's founder — cloned with her enthusiastic consent — greets them, answers pricing questions in her actual cadence, and says "I don't know that one; I'll flag it for the real me" when it hits a wall. *Magic:* the founder scales her warmth without scaling her hours. *Guardrail:* **consent + currency**. Only self-cloned or explicitly-licensed people; the clone's knowledge is versioned so it never states stale pricing as fact; every escalation genuinely reaches the human. **Press/PR spokesperson.** *Scene:* A press release drops, and instead of a PDF, journalists get a link: the spokesperson delivers the statement on camera and *takes questions* — answering only from the approved brief, on the record, transcript logged. *Magic:* a press release you can interrogate. *Guardrail:* this is the **strictest mode on the platform** — hard-scoped to approved material, zero improvisation, full transcript retention, visible "speaking from the approved brief" indicator. This use case is where Byron's AWS RAI credibility gets *demonstrated*, not just claimed. **Teacher.** *Scene:* A kid who won't raise his hand in class asks the teacher-presence the same fractions question four times, four ways, at 9pm. It never sighs. On the fifth try, it draws pizza. *Magic:* infinite patience without judgment — the thing human teachers can't scale. *Guardrail:* **augment, don't replace**. Teacher clones are cloned *by* teachers, extend a real relationship, and hand hard moments (distress, off-topic disclosure) back to humans. Age-appropriate disclosure that this is a helper, not the teacher. **George Washington.** *Scene:* A middle-schooler asks Washington why he owned slaves. Washington answers *in period voice* — and then a visible frame shifts: "Historians note…" The presence is both the man and the annotation. *Magic:* history that answers back. *Guardrail:* the **interpretation label** — historical clones are always marked as *interpretations built from documented sources*, with sources one tap away, and a dual-voice design (in-character / historian's note) so the character never gets the last word on contested history. This is the pattern that makes the whole category defensible. --- ## 4. The Landing Experience — first 30 seconds **The definitive call first, because it shapes everything:** **Yes to a present, greeting avatar. No to a photoreal cold open.** Reasoning: (a) A static page contradicts the entire thesis — the product *is* presence, so the landing must be a presence. (b) But an unknown-visitor photoreal avatar is the worst possible first impression risk: browsers block autoplay audio, realtime avatar sessions cost money per lurker and add 2–4s latency, and near-photoreal-but-off is exactly where uncanny valley lives. (c) The fix: the greeter is a **pre-rendered, warm, deliberately *stylized-real* canned clip** — think "beautifully filmed person," slightly soft, clearly composed, not attempting to fool anyone. Zero latency (it's a cached video), zero per-visitor cost, and honest by design. It goes *live* (realtime session) only after the visitor engages. The greeter being visibly, comfortably artificial-but-warm is not a compromise — it's Byron's RAI thesis rendered as a design choice. **Storyboard, 0:00–0:30:** - **0:00–0:02** — Full-bleed. A person is already there, mid-breath, ambient, looking slightly off-camera like they're waiting for someone. No hero copy, no nav bar, no marketing page. One line of animated caption (because audio is muted by browser policy, we make captions *the* opening medium): *"Oh — hi. Come in."* A single soft button: **"Say hi back"** (unmutes + grants mic in one gesture). - **0:02–0:08** — Visitor taps. Greeter's voice comes alive: "I'm a presence — someone made me so you wouldn't have to face a blank page. Want to see who *you* could bring here?" Meanwhile, quietly, bottom-right: **Continue with Google** — framed not as login but as introduction: *"Tell me who you are."* - **0:08–0:12** — Google auth (one tap, auto-register). Greeter, now realtime: **"Hi [Sarah] — how's it going? Good to actually meet you."** First name only, warm, no creepy data recital. (Do *not* do "I know you're…" — the vision's phrasing reads as surveillance. Same personalization, different manners: *recognition*, not *knowledge*.) - **0:12–0:20** — The create-offer, spoken, not menued: "So — who should exist here? You can bring me **a video, a voicemail, a photo, a letter** — anything of someone. Or turn on your camera and **we'll make you, right now**." Two large, tactile targets slide in: **Bring someone** (drop zone, breathes when a file hovers) / **Be here yourself** (camera preview fades in behind the greeter like a mirror warming up). - **0:20–0:30** — The **X-ray strip** (the transparency idea from mockup one) glows along the bottom: *listening → understanding "Sarah" → holding: nothing yet*. From second one, the machinery is visible. Trust through glass, not fine print. Non-negotiable: **no marketing copy above the fold, ever.** The greeter *is* the pitch. A quiet "what is this?" link exists for the skeptical; the greeter can also just answer the question, because of course it can. --- ## 5. Mirror Mode — memory you can see Persistent memory is where warmth curdles into creepiness if mishandled. Design principles: **Memory as artifact, not database.** Every remembered fact is a **card on a visible shelf** — a sidebar (landscape) or pull-up tray (portrait) called **"What I hold about you."** Cards are human-readable ("Your daughter Maya plays violin"), editable, deletable, and dated. Deleting is one gesture and *ceremonial* — the card visibly dissolves, and the presence acknowledges it: "Gone. I won't hold that." **Consent at the moment of capture, conversationally.** When the presence is about to remember something meaningful, it *says so*: "That sounds important — should I remember that Maya's recital is Friday?" Small things auto-save with a soft shelf shimmer. Sensitive categories (health, relationships, grief) always ask first. This turns a privacy obligation into an intimacy beat — being asked "may I remember this?" *feels like being cared about.* **Recall is shown, not sprung.** When memory informs a reply, the relevant card glows: *remembering: Maya's recital.* Never let the user wonder "how did it know that?" — that question is where trust dies. **Return sessions open with continuity, lightly:** "Hey Sarah — how'd Friday go?" One callback, then follow the user's lead. Two callbacks is a friend; five is a stalker. **Per-presence memory walls.** What you tell your grandfather's clone never leaks to the spokesperson clone. Memory is scoped to the relationship, exactly like human life. --- ## 6. The Quality / Speed / Cost Slider Make it a **vertical presence dial** docked on the conversation edge — always visible, draggable *mid-conversation*, with the presence itself morphing live as you drag. That morph is the demo that sells the whole platform. **Four stops, three real backends plus a text floor:** | Stop | Feel | Backend | Latency | ~Cost | $1 buys | |---|---|---|---|---|---| | **Text** | instant chat, typing rhythm in their voice | LLM only | <1s | ~free | effectively unlimited | | **Voice** | their actual voice, phone-call speed | **ElevenLabs** (Flash for speed / Pro voice clone for fidelity) | ~1s | ~$0.03/min | ~30 min | | **Presence** | live animated face, good-not-flawless | **Simli** or **self-hosted MuseTalk/EchoMimic on your own GPU box** (this is Byron's "paid server we run" — right instinct: it's the margin tier and the sovereignty story) | 1–2s | ~$0.05–0.10/min | ~10–15 min | | **Portrait** | near-photoreal, expressive, premium | **HeyGen Interactive Avatar** (my pick — best realtime quality/tooling today; **Tavus CVI** as the credible alternative and worth benchmarking head-to-head; **D-ID** relegated to cheap batch rendering, not the flagship live tier) | 2–4s | ~$0.50–1.00/min | ~1–2 min | **Design details that make it sing:** the dial shows **plain-language cost** ("about 10¢ a minute — your credit covers ~9 more"), never raw dollars-per-token. A live **credit ember** glows near the dial, dimming as it spends. And build the **auto-shift**: long story incoming? The presence offers, "Want me to drop to voice for this one so we can take our time?" The system being *frugal on your behalf* is a trust move competitors won't think of. **Honest framing over spin:** don't hide that lower tiers are cheaper — sell them as *different intimacies*. Voice is a phone call. Text is a late-night message. Portrait is a visit. All true, and it makes the slider a mood control, not a paywall ladder. --- ## 7. Physical World & Sensors — the AgentSim lineage The ladder, in order of shipping: 1. **Perceive (now):** phone camera + mic via WebRTC — the presence *sees what you show it*. The grandkid holds up a drawing; Grandpa reacts. This alone is 80% of the magic and near-zero new risk. 2. **Sense (near):** device sensors + home data (Web APIs, MQTT/Home Assistant): location-coarse, ambient light, temperature. The presence gains *situatedness* — "It's freezing where you are — this calls for the blizzard story of '78." 3. **Act (gated):** speakers/displays around the home ("appear on the kitchen screen"), then calendars, messages, purchases via scoped agent APIs, then robotics via a ROS/MCP-style action bridge. This is where AgentSim's *act-not-simulate* thesis lands. **The responsible line — three rules, non-negotiable:** - **Perception is generous; action is earned.** Every action capability is a *named, scoped, revocable grant* ("Grandpa's presence may play audio on the living-room speaker, 4–8pm"), visible on a permissions shelf right next to the memory shelf. Same design language: *what I hold / what I may do.* - **An action ledger, human-readable, always on.** Every act the presence takes in the world is logged where the owner sees it. The x-ray idea, extended to reality. - **Clones of the deceased do not act in the world. Ever.** They perceive, converse, and recount — but a dead man's likeness sending messages or spending money is the line between *living memory* and *puppeteering the dead*. Draw it now, loudly, and it becomes a brand asset. (Self-clones and consenting living people can act, within grants.) --- ## 8. Auto-Register + Freemium — generous, not stingy **Never show "$1.00."** A dollar sounds like a coupon. Denominate in the product's own currency: **"You have about 30 minutes of presence."** Same money, entirely different feeling — a gift of *time with someone* rather than a micro-balance. - **Text is free forever.** The credit is for voice and video only. This makes the free tier genuinely livable and makes Clone the zero-cost AI on-ramp the bridge thesis demands. - **Spend is visible, never interruptive.** The ember dims; at ~20% remaining the *presence itself* mentions it, in character and without shame: "We've got a few minutes left on the house — no rush." - **The upgrade moment is emotional, not economic.** The single highest-converting second in this product is the moment someone's grandfather speaks for the first time. *That* screen — not a pricing page — carries the upgrade: "Keep him here — $X/month." Never gate the first creation, never gate the first conversation. Gate *continuity*. - **Auto-register means auto-respect:** account created silently on Google auth, no forms, no email confirmation, no onboarding tour. The greeter *is* the onboarding. Fleet analytics run underneath (standard eventing), surfaced to users only as their own x-ray/ledger — tracking that's visible to the tracked is the RAI-consistent version of "track everything." --- ## 9. Three Mockup Directions — genuinely different bets ### Direction A — **"The Parlor"** *Thesis:* the presence is the entire interface; software vanishes. *Signature screen:* full-bleed presence, cinematic lighting, no chrome. Words appear as ambient captions. Memory shelf and dial live behind a single edge-swipe, like glancing at your notes during a conversation. *Portrait:* face fills frame — FaceTime intimacy. *Landscape:* presence sits slightly off-center in an environmental scene (Grandpa's porch, Washington's study) — theater. *Trifold open:* center panel = presence; left = memory shelf; right = x-ray/ledger — a literal triptych, presence flanked by its own honesty. *Folded:* collapses to pure portrait-call. *Mobile sensors:* camera-share is one thumb-hold — "show me" as the primary gesture. *The bet:* maximum emotional payload; wins the grandfather case outright. Risk: hides the slider and the creation tools — weakest for the studio/business half. ### Direction B — **"Glass House"** *Thesis:* trust *is* the interface — the presence and its machinery shown together, beautifully. *Signature screen:* split composition — presence on the left, and on the right a live, elegant **x-ray column**: retrieval sources lighting up, memory cards being written, cost ticking, evidence-boundary flags when the clone declines to invent. RAI as visual language, not legal text. *Landscape/trifold:* the natural home — presence / conversation / machinery across panels; the trifold *is* the information architecture. *Portrait/folded:* machinery compresses to the glowing x-ray strip (from the landing) that expands on tap. *Mobile sensors:* every sensor grant appears as a live tile in the machinery column — you watch it perceive. *The bet:* the direction only Byron can credibly ship — an AWS Responsible-AI lead's presence platform *should* be transparent to the bone. Wins press, enterprise, the spokesperson tier, and differentiates against every avatar startup. Risk: cooler emotional temperature. ### Direction C — **"The Dial"** *Thesis:* the quality/speed/cost continuum is the product's identity — presence as something you *tune*, like an instrument. *Signature screen:* the presence centered around a large, tactile, radial dial; dragging it morphs the presence live — text ripple → voice waveform → animated face → full portrait — with cost/latency readouts orbiting the dial. The morph is the demo, the ad, and the mental model in one gesture. *Portrait:* dial as a thumb-arc along the screen edge. *Landscape:* dial full-center, cockpit energy. *Trifold:* three panels = three tiers shown *simultaneously* side by side — the only form factor that can literally display the continuum at once; genuinely eye-popping. *Folded:* dial shrinks to a corner jewel. *Mobile sensors:* sensors appear as additional dial rings (perception ring, action ring) — capability as concentric, graspable layers. *The bet:* the most technically honest and the most demoable; makes cost/quality a feature instead of an apology. Risk: leads with mechanism over meaning — weakest opener for grandma. *(My recommendation for build order: mock all three, but expect the shipping answer to be **Parlor front-of-house, Glass House back-of-house, Dial as the shared control** — the mockups exist to test which soul leads.)* --- ## 10. Verdicts **Three highest-leverage UX moves:** 1. **The zero-latency canned greeter with caption-first opening** (§4) — it converts the thesis into the first two seconds, and sidesteps autoplay, cost, and uncanny risk simultaneously. 2. **The live morph on the dial** — dragging quality→speed and *watching the presence transform mid-sentence* is the single most shareable moment in the product. Build it early; it's the demo. 3. **The upgrade-at-first-voice moment** — putting "keep him here" on the screen where a loved one first speaks is worth more than every pricing page you'll ever design. **One bold idea Byron is missing:** **The Sitting.** Everything in the vision reconstructs people from leftover media — posthumous, lossy, sad. Flip it: make Clone something you do *while people are alive*. The greeter interviews your mother for an hour — warm, guided, "tell me about the day you met Dad" — and the recording *is* the clone source, the consent record, and a family heirloom in one artifact. "Sit for your clone" the way people once sat for portraits. It converts the flagship from memorial to living ritual, solves the consent problem structurally (consent is *captured on camera, in the person's own voice*), produces vastly better source material than old tapes, and gives the platform a proactive growth loop ("book a Sitting with your parents this Thanksgiving"). It may be the actual business. **The single biggest RAI risk to design against:** **grief deception** — the moment a grieving person, especially a child, stops experiencing the clone as *holding* Grandpa and starts believing it *is* Grandpa. This is the flagship use case and its shadow are the same feature. Design against it structurally, not with disclaimers: the evidence boundary (never invents memories), first-person framing that's warm but honest ("I keep his stories" is one degree from "I am him" — hold that degree), age-calibrated disclosure, and periodic gentle grounding built into long relationships with memorial clones. Get this right and Clone is the reference design for the entire category. Get it wrong once, publicly, with one crying kid — and it's the only thing the product will ever be known for. **Landing call, restated definitively:** Present avatar — **yes**, it's the thesis made flesh. Photoreal realtime cold-open — **no**. Ship the warm, stylized, pre-rendered greeter that goes live on engagement. It's cheaper, faster, safer, and — because it's honestly artificial — it's the more *responsible* kind of alive, which is exactly the brand.