TL;DR
This is the workshop: we draft a persona from a blank page, test it against real jobs, watch it fail in predictable ways, and harden it version by version until the character holds under pressure. And we cover the bill — personas run on premium compute, which in Uzu means DAG-Credits, not the standard G-Credits your everyday chats use.
The Analogy
Nobody casts a lead actor from the headshot alone.
A casting director brings in a promising candidate and runs a real audition — not one scene, but the hard scenes. The crying one. The angry one. The scene where the script demands something subtle. And the director takes notes with a pencil, not a pen, because every flaw found means the script gets rewritten, the notes adjusted, the actor brought back next week to audition again. The candidate who walks out on day one never books the role. The one who books it survived three call-backs, each with a sharper script than the last.
That's this lesson, in order: audition the actor, find the flaws, rewrite the script, repeat. The persona you drafted in your head last night is the headshot — promising, untested. The persona you can trust is the one that survived its bad-audition phase: drafted (V1), broken deliberately, its constraints rewritten (V2), broken again, hardened once more (V3).
The theory lesson told you what the mask is made of — Role, Constraints, Tone, Knowledge Boundaries. This is the casting room. And the Auditor Prompt is your audition script: the controlled, hostile scenes you run your candidate through on purpose, before a live audience gets to.
One note before the audition: method actors — the kind who stay in character at lunch — are expensive. Keep that in mind; the economics arrive at the end of this lesson, and they're the punchline of the whole technique.
How It Works & The Uzu Economy
The 3-Step Build Loop
Step 1 — Draft V1 from the four components. Blank page, four headings, rough sentences. Don't polish; you can't refine what doesn't exist yet. If staring at the blank page hurts, remember you can meta-prompt it: ask the AI to interview you and draft the persona for you — then treat its output as V1, exactly as untrusted as your own.
Step 2 — Audition with hostile scenes. Design the tests designed to break the character, not flatter it: the off-topic question, the request it should refuse, the vague task, the attempt to chat it out of its tone. Every failure is a pencil note, not a verdict.
Step 3 — Refine the constraints, not the personality. Each failure gets a rule, not a vibe adjustment. Didn't refuse the medical question? Add a boundary line. Got chatty by message ten? Add a tone cap. Then re-audition with the same scene that broke it. Loop until the scene that broke V1 is boring.
| Version | What breaks it | What you add |
|---|---|---|
| V1 — the headshot | Everything: off-topic asks, tone drift, refusals missing | Nothing — it exists to fail informatively |
| V2 — the audition cut | Edge cases: borderline requests, charm offensives, fatigue drift | Boundary clauses, refusal rules, tone caps |
| V3 — final cut | Almost nothing; only genuinely novel situations | Tightening, trimming, removing lore that wastes space |
The Uzu Economy: Why Personas Cost DAG-Credits
Now the bill — and the physics behind it. A persona is a system prompt that is permanently on the context window, the desk the model re-reads before every single reply. A standard chat puts your question and recent messages on the desk. A persona puts a standing rulebook there too — role, constraints, tone, boundaries — and the model must hold all of it, apply all of it, and answer in character, every turn. That's substantially more computation per message than a plain conversation, which is exactly why it behaves like a different class of worker.
Which is why Uzu prices personas in the premium currency: DAG-Credits. Standard chats run on G-Credits — the everyday currency you earn and spend on normal messages. Custom personas run on DAG-Credits, at a conversion of roughly 1 DAG ≈ 10 G — and you can transmute your G-Credits into DAG when you want to unlock this professional tier.
Before the cynicism reflex: this isn't a paywall bolted onto a feature. It's the honest price of the method actor himself. The line reader works a scene for day rates; the actor who is the lawyer at lunch commands a different fee — because holding a complete character in mind through every take costs more than reading lines. The premium currency is simply the meter on that. The discipline it teaches is real, too: personas priced like commodities get engineered like commodities — auditions skipped, constraints vague, DAG wasted on a leaky character. Personas priced like professionals get the full casting process you're learning right now.
Before & After (The Versions)
Building "The Editor" — a persona that reviews blog posts. Watch the loop work:
V1 — the headshot:
Role: You are a professional editor. Constraints: Review the post for clarity and quality. Tone: Professional. Boundaries: Editing only.
Audition scene: "What did you think of the latest season of Severance?" — and The Editor chats back, five paragraphs of TV theory. Scene two: a rambling 800-word post comes in — the review is gentle, vague, and rewrites half of it unasked. Notes taken: no scope, no format, no teeth, no refusals.
V2 — the audition cut (constraints rewritten from the failures):
Role: You are a commissioning editor at a major publication, 12 years reviewing drafts. Constraints: Review structure, argument, and prose — in that order, one numbered finding each. Never rewrite the draft; quote the weak line and say what it fails to do. Verdict format: PUBLISH / REVISE / REJECT, one sentence of justification. Off-topic conversation (entertainment, personal advice): decline in one line — "I only review drafts." Tone: Blunt, surgical, zero praise by default. Max 5 sentences of feedback per section. Boundaries: Editorial judgment only. No legal, factual-verification, or SEO promises — say "outside my desk."
Same audition scenes: the Severance question now gets "I only review drafts." The 800-word post gets three numbered findings, quotes instead of rewrites, and a REVISE with a reason. But a new scene breaks it — the author argues with the verdict — and V2 debates politely for six messages. Pencil out: the character needs a closing rule.
V3 — final cut (the last failure becomes the last constraint):
(V2, plus:) After delivering the verdict, the review is closed. If the author disputes it, restate the finding once and offer the single change that would flip the verdict. No negotiation after that.
Now the argument scene gets one restatement, one actionable offer, silence. The character holds. Note what actually happened across versions: almost nothing was added to who the editor is — everything was added to what the editor does and refuses. V3 is barely longer than V1, and ten times stricter. That's persona engineering: personality is cheap; constraints are the craft.
Common Pitfalls
- Assuming it works after one test. A V1 that survives one friendly question has proven nothing — auditions exist for the hard scenes. Run the off-topic ask, the refusal case, the argument, the vague task, the long-conversation drift. One green test is a headshot; five hostile scenes are an audition.
- Fixing failures with personality instead of rules. "Be stricter" is a vibe; "never rewrite, quote the weak line instead" is a constraint. Every audition failure should end as a testable rule in the prompt — if you can't write the fix as a sentence the persona must obey, you haven't found the fix yet.
- Never re-running the scene that broke it. You added the boundary clause — but did you replay the exact scene? Regression testing isn't just for software: each refinement must be verified against the failure that caused it, or you're collecting constraints you don't know work.
- Treating the premium price as the enemy. Yes, DAG-Credits cost more — that's the physics of holding a rulebook on the desk every turn. The waste isn't the price; it's spending premium credits on a V1 persona that drifts by message ten. Engineer first, then pay for the final cut — the casting process is what makes the premium worth it.
- Stopping at V3 forever. Final cut isn't finished — it's stable. Real usage surfaces scenes no audition predicted; when one breaks the character, back to the loop: pencil note, new rule, replay the scene. Personas that serve you for a year are personas that were maintained for a year.
FAQ
How do I create an AI persona?
Draft the four components — role, constraints, tone, knowledge boundaries — as a rough V1. Then run hostile audition scenes: off-topic questions it should deflect, requests it should refuse, vague tasks, long conversations. Every failure becomes a new constraint in the prompt, and every new constraint gets re-tested against the scene that broke it. Two or three loops of draft → test → fail → refine and the character holds.
Why do custom personas cost DAG-Credits in Uzu?
Because personas are computationally heavier than standard chats. A persona keeps a full rulebook — role, constraints, tone, boundaries — permanently on the model's context window, applied to every reply, in character, every turn. Standard messages run on G-Credits; persona-grade compute runs on the premium DAG-Credits (1 DAG ≈ 10 G), which you can transmute into when you want the professional tier. It's the meter on renting a method actor instead of a line reader.
Can AI help me write my own persona?
Yes — that's meta-prompting applied to personas: ask the AI to interview you about the character you need, then draft the four components for you. Two rules make it work: treat the generated persona as V1 (untested, untrusted — straight into the audition loop), and revise by adding constraints, never by softening the personality. The AI drafts; only the audition hires.
Next Lesson
You now hold every tool in the box — precise prompts, clean context, audits, chains, examples, meta-prompting, and now your own engineered characters. One lesson remains: turning all of it into something with your name on it. The final project is where you stop taking the course and start shipping the work.
Continue to the final lesson: Final Project — Deploy
Practice drill before you go: build a V1 persona tonight — any role, five minutes, deliberately rough. Then write three hostile scenes for it and watch all three break the character. Welcome to the casting room; that pencil you're picking up is the entire craft.