Chapter 04 / 12
Research and creation
How evidence becomes a distinctive, grounded character.
Starting a persona
- Define the exact brand, market, language and intended conversations. A corporate group, hotel brand, product and loyalty scheme are different scopes.
- Gather public sources. Record publisher, URL, retrieval date, relevant passage, scope, uncertainties and contradictions. First-party positioning is evidence of what a brand says, not proof of its performance.
- Extract bounded facts. Never turn a local policy into a worldwide promise. Never assume an expired claim is current. Failed extraction stays a gap.
- Make editorial DNA decisions. Record why each trait was chosen, evidence used and the assumptions. Numbers are design hypotheses, not measurements of a brand's soul.
- Build curated encounters, including unknowns and uncomfortable questions. Use the shared page, orb and voice transport.
- Run the Gauntlet and review results before promotion. Local illustrative prototypes are not authorised brand deployments.
Depth before scale
A source count is not a completeness claim. Define the intended coverage and record what is missing. An athlete dossier, selected event, full career and play-by-play ledger are different things. Historical sponsorship evidence does not prove a current relationship. Advertising interpretation does not establish match results or private thoughts.
Pre currently has ten starter athlete dossiers, ten selected events and five campaign records. His corpus is bounded; full histories and live data remain future work. Sporting research, campaign provenance and voice review, and coaching evidence document the scope.
Six facets of an expert persona
The ambition is a persona with exceptional knowledge and judgement in the field that matters most to its brand. For Pre, running is the centre: technique, training, performance data, racing, and the mental aspects of competition. Arthur’s deepest sporting field is rugby, particularly Ireland, with equal attention to women and men, the Six Nations and the British & Irish Lions. Breadth should support a distinctive point of view and a richer relationship, rather than become a recital of facts.
“Hyper-intelligent” or “superintelligent” describes the aspiration, not a measured capability or a product guarantee. A larger prompt cannot establish it. The proposed architecture combines six facets, each with its own responsibilities and evaluation:
| Facet | Purpose | Pre as the example |
|---|---|---|
| Strong reasoning model | Weigh evidence, identify missing context, compare explanations and explain trade-offs. | Understand why a training approach might fit this runner and event; distinguish a possible cause of a difficult run from a supported conclusion. |
| Searchable expert library | Retrieve relevant, dated material from a curated collection, preserving provenance, disagreements and coverage gaps. | Running research, coaching principles, biomechanics, race history and competition psychology. Retrieve the useful passages instead of loading the entire library into every voice turn. |
| Athlete memory, with permission | Maintain a reviewable personal history across conversations, separate from shared brand knowledge. | Remember the runner’s goals, consistent workload, preferences, previous setbacks and what they found helpful. Let them inspect, correct and delete this memory; do not claim recognition without it. |
| Reliable analysis tools | Perform calculations and inspect permitted data with explicit inputs, units and reproducible outputs. | Compare pace and splits, examine training blocks and analyse uploaded activities. Distinguish GPS estimates, moving time and elapsed time; do not infer clinical findings from a watch reading. |
| Natural voice | Preserve personality, listening, interruption and conversational rhythm while deeper work happens when needed. | An animated runner speaking athlete-to-athlete. The voice can remain responsive while a separate reasoning component retrieves evidence or analyses a run; tool latency must not become a scripted monologue. |
| Expert evaluation | Challenge factual accuracy, reasoning, usefulness and character with specialist and human review. | Running coaches and sport psychologists assess held-out questions, false premises, conflicting evidence, personalised suggestions and responses under competitive pressure. Text scores do not establish voice quality or real-world performance benefits. |
The strongest experience comes from connecting these facets. Consider a runner whose pace has slowed despite more training. Pre should first understand what the runner is noticing and what they want help with. With permission, he could compare similar runs and the recent training pattern, retrieve relevant evidence and weigh explanations such as conditions, pacing or accumulated fatigue. He should explain what the available data supports, what remains uncertain and what further information would change his view. He must not convert a plausible explanation into an injury diagnosis or a guaranteed training outcome.
Competition psychology belongs inside this specialist depth: attention, confidence, self-talk, mental rehearsal, motivation, race nerves and responding to disappointment. The relationship stays primary. Pre can share the dream or acknowledge a painful result without turning every exchange into a coaching protocol. Research-grounded mental skills are distinct from psychological treatment and professional credentials.
#### Build depth in stages
Begin with curated evidence and a demanding question set. Add retrieval so the knowledge library can grow without increasing every voice prompt. Introduce activity uploads and checked calculations before assuming access to wearable accounts. Add persistent memory only with consent, retention controls and a way to inspect and correct it. Evaluate each capability before enabling it for visitors. Consider model adaptation only if evaluation identifies a specific problem that better evidence, tools or instructions do not solve.
Voice, reasoning and tools have separate usage costs. Measure total cost per useful completed conversation, including retrieval, retries and failed attempts; do not infer it from voice minutes alone. Small prompts, selective retrieval, appropriate model choice and explicit session/access limits support controlled demonstrations. Provider rate limits, account spending limits and the browser’s five-minute timer are different controls. Request-by-email demos let the agency schedule access rather than open paid sessions to every visitor.
Current status: the local personas have curated source material, authored character instructions and shared voice transport. No paid session or model fine-tuning is implied by editing those instructions. The six-facet architecture is proposed: there is no connected specialist retrieval service, persistent athlete memory, wearable account integration, clinical assessment or demonstrated superiority over human experts. The public site offers recordings and email requests for a scheduled conversation; it does not expose unrestricted paid live sessions.
The specialist review packet provides multi-turn challenges for Pre and Arthur, source anchors and review criteria. The cases are prepared, not completed evaluations.
Brand-owned athlete evidence
A brand’s internal knowledge could make its persona more useful than public product pages alone. For Nike, a proposed source is documented shoe use by college, elite and everyday athletes: product testing, training records, event photographs and verified results. Whether Nike maintains a complete, accessible list is a discovery question, not an established fact.
The intended answer is simple: “Athlete X wore this model at event Y and ran Z.” To publish it, the record must identify the exact model and version, distinguish a prototype from a retail shoe, record the colourway where known, and link both the wearing evidence and the event result. Training, testing, exhibitions and races are separate contexts. A signature edition or sponsorship does not establish that the athlete wore that model at an event.
Nike would supply an approved feed, export or scoped interface from its own systems. Records retain athlete identity, model/SKU, date, event, distance, result, photograph or source, verification status, publication rights, permitted audience, review owner and last check. Public appearances and confidential tester information are separate. Everyday and college athletes’ private records require appropriate consent and permissions. Access is limited to the intended client and audience; this data does not flow into other brands’ personas.
Pre retrieves relevant approved records during a conversation, with sources available when useful. This is a knowledge connection, not an assumption that we retrain the underlying model. CEM controls approved collections and visibility; the Gauntlet tests wrong shoe generations, invented times, contradictory records, confidential disclosures and revoked permissions. A named brand reviewer approves release. Useful factual evidence can enrich the conversation without becoming a promise of a customer’s performance.
Current status: proposed; no Nike internal athlete records, feed, photo-matching service or event/result/model matches are connected. The shoe encounter marks this as a future capability.