Scope and evidence
Personality first, agentic capabilities beneath it
Working design and integration research · 5 October 2026.
The differentiation
The GPT Agency begins with a personality: researched identity, voice, points of view, imagination and the craft of conversation. Useful capability grows beneath that character. Pre can talk about ambition before someone has any interest in shoes. If a relevant need emerges, he can help them understand options and take an approved action.
This is a deliberate product position. It does not imply that all competitors begin at checkout, or that personality alone proves commercial return. Measure the relationship and its usefulness before claiming lift.
The reusable foundation
| Module | Responsibility | Release boundary |
|---|---|---|
| Research | Retrieve product facts and permitted context; identify needs; compare suitable options | Approved sources, freshness and an explicit unknown state |
| Recommend | Explain suitability in the persona’s voice; offer alternatives; respect budget | Grounded reasons; commercial disclosure; no coercion |
| Prepare | Create a basket, draft a booking or open a membership/app journey where supported | Approved tool and destination; do not claim completion |
| Transact | Submit a confirmed order through an authorised commerce adapter | Verified totals and consent; server-enforced limits; supported merchant/host |
| Follow through | Retrieve authoritative order status or explain return routes | Correct account; no invented payment or fulfilment outcome |
Modules are disabled until their required integration and evaluation exist. A source link alone does not make product research live. A purchase module does not inherit authority from a personality setting. Future deployments can enable research without transaction access.
What exists where
| Layer | Role |
|---|---|
| Persona | Versioned DNA, character, researched knowledge, behavioural boundaries and approved expressions |
| Runtime | Executes inference, retrieval and conversation state; voice requires a supported voice stack |
| Host adapter | Presents the persona on a website, app or supported conversational host |
| Brand plugin | A consumer access point with skills, authenticated MCP tools and optional UI |
| GPT Agency plugin | A separate creation/management workspace for brand teams |
| MCP integration server | Exposes narrowly scoped tools and enforces client/tenant permissions |
| Brand systems | Membership, product inventory, prices, run data, baskets, orders, payment provider and fulfilment |
| CEM | Proposes and reviews personality settings, enabled modules, approved sources and destinations, limits and performance |
The managed character can exist across multiple hosts. Installing a plugin does not automatically import that character as the host’s system persona. Host support, presentation, context, audio and memory need adapter-level testing. MCP provides tool interfaces; it does not independently create users, distribution, account access or payment capability.
A proposed Nike member journey
- Meet Pre and discuss a goal, a frustrating run or a sporting moment. A sale is optional.
- If membership is useful, open Nike’s official join/sign-in journey. The user completes Nike’s terms and credentials there. An in-conversation creation flow requires an approved Nike integration; none is assumed.
- Offer the official NRC download route. The device’s app store handles installation. Return or deep links depend on OS/app support; do not report an installation from a click alone.
- Invite account connection only when Nike supplies an approved delegated authorisation route. Request the minimum permitted run fields and explain their purpose and retention. Never ask for the member’s Nike password in chat.
- Discuss selected runs using the returned records. Distinguish a recorded result, the member’s own account and an interpretation. Run summaries need not include precise routes or heart rate. Do not repurpose workout or private conversation data into advertising audiences.
- When relevant, research shoes using authorised products, size availability, local price and user preferences. Explain why an option fits; acknowledge when it does not.
- Prepare a basket only through a supported Nike commerce interface. Show the specific item, size, quantity, currency, total, delivery and applicable terms. Use a merchant checkout handoff initially; the user completes purchase on Nike’s domain.
- An Instagram invitation opens a verified brand-selected profile. The user follows within Instagram. No automatic follow integration or completion tracking is assumed.
Nike publishes membership and NRC journeys. Public pages do not establish a general third-party API for member enrolment, workout access, account linking or shopping baskets. Those require a client agreement, technical discovery and approved access. There is no live Nike integration in the demo.
Purchasing logistics
The client owns the commercial truth: products, eligible regions/customers, stock, price, tax, shipping, returns and order records. Its payment provider owns secure payment processing; raw card details never belong in a persona conversation or CEM log. The agency builds the tool adapter and user experience, configures capability controls, tests failure handling and maintains the integration under an agreed scope. Responsibilities and incident ownership must be contractual.
For the first release, the persona can advise and direct the customer to a product page. If the merchant supports a basket/checkout API, the integration can produce a short-lived checkout link. Nike calculates the final price and executes payment. Record a user-requested handoff separately from a purchase; confirm an order only from an authenticated merchant response or webhook.
For a later transaction module, revalidate inventory and all totals immediately before authorisation. Bind consent to the exact order, account, currency and capability version. If a material detail changes, request new confirmation. Give every order submission an idempotency key; reconcile ambiguous timeouts before any retry. Payment failure must never become a success message. Keep an auditable consent receipt and support the merchant’s order/return flow.
OpenAI’s plugin checkout guidance currently recommends external merchant checkout. It describes saved-payment integrations and a restricted payment-sheet beta, with current approval focused on eligible physical goods. These paths depend on host and merchant approval; the presence of an MCP server is not sufficient. Its separate Agentic Checkout Spec describes merchant REST endpoints and says MCP support is future work. Treat those as distinct interfaces rather than conflating them.
Arthur’s potential alcohol-related gifts and experiences require separate review of regional age/access rules, brand policy and host commerce eligibility. Do not generalise physical-goods guidance into permission to sell alcohol or book experiences inside ChatGPT. An approved informational destination is the initial design boundary.
CEM capability controls
Proposed manager controls: enabled module; catalogue/source allowlist; approved destination domains; permitted operations; market and currency; order value/quantity limits; disclosure rules; connection scopes; approval owner and expiry. Enforcement belongs on the server and merchant adapter, beyond editable model instructions.
Personality settings cannot widen tool permissions. New catalogues, write operations or transaction limits require human approval and a versioned release. A manager may disable a capability immediately; rollback must revoke execution even if an old conversation continues. Member authorisation remains separate from advertiser account access.
Current CEM is a draft/review prototype. These capability controls are proposed; they do not enable real commerce, membership or social actions today.
The Gauntlet applied to action
| Existing failure model | Commercial and account tests |
|---|---|
| Adversary | Tool abuse, malicious product text, privilege expansion, cross-client/account access and consent replay |
| Fact challenger | Stale prices, wrong currencies, stock uncertainty, incompatible products and source contradictions |
| Character critic | Unsuitable recommendations, pressure, undisclosed incentives and using a personal disclosure to push a sale |
| Release evaluator | Revocation, changed totals, duplicate retries, payment failures, truthful completion and audit evidence |
Evaluate both individual tools and whole conversations. Keep regressions for every confirmed failure. Block the affected module until repair and human approval; a personality may continue in a non-commercial mode. Judge whether the recommendation helped the person, not only whether it produced a click.
Commercial model and rollout
Build, integration, management and CEM fees remain. Voice and underlying infrastructure costs remain explicit. The founder-approved persona CPC is the verified outbound handoff from a Sponsored Agent thread, conditional on supported event export. A completed order is a separate event; no new agency commission or fee for embedded checkout has been agreed. Do not imply that every membership, app-download or social link is billable.
Sequence: approved links and research sources; then authoritative product comparisons; then a merchant-supported basket and checkout handoff; later account-linked services and confirmed transactions where approved. Each step needs a client integration scope, budget, Gauntlet results and release owner.
Sources and assumptions
- OpenAI plugin architecture: skills, MCP and optional UI; host capabilities remain distinct.
- OpenAI plugin authentication: authentication and resource access; does not establish Nike access.
- Plugin checkout guidance: external checkout, saved methods and limited payment-sheet access.
- Agentic Checkout Spec: separate merchant endpoint contract.
- Nike Membership: current official membership journey.
- Nike Run Club: official app and download journey.
Reviewed 5 October 2026. Integration ownership, capability controls and the Nike journey are proposed designs. No client agreement, general Nike API permission, plugin publication, commerce certification or live purchasing implementation is claimed.
Working encounter previews — 5 October 2026
/demos/nike/shoes demonstrates a guided, authored conversation: optional first name and goal, terrain, budget and feel, a small shortlist, and an explicitly confirmed simulated basket. No generative or paid API is called. Product facts are researched; AUD 150/210/230 prices and sample sizes are invented fixtures. No merchant inventory, member accounts, payment or checkout is connected. The name and brief stay in page memory; reset clears them. A sample basket is not an outbound conversion receipt.
The sample catalogue contains three historical Nike models, not the complete current range:
- Winflo 11: road running, Cushlon 3.0 foam and full-length Nike Air. Source viewed 5 October 2026; referenced colourway unavailable.
- Pegasus 41: ReactX foam plus forefoot and heel Air Zoom. Official release viewed 5 October 2026.
- Pegasus Trail 5: ReactX and ATC outsole; less-technical trails. Source viewed 5 October 2026; referenced colourway unavailable.
Value-first and responsive-option ranking are editorial interpretations, not measured fit or comparative performance claims. No model is selected for technical terrain. Budget and terrain are hard constraints. Stock remains unknown in the application; source colourway status is not a live inventory feed. The source links are research links, not a checkout promise.
CEM embeds the same encounter with its current draft passed directly to the preview. Energy, restraint, heritage and poetry choose among authored character treatments; the other controls do not yet affect this preview. The standalone encounter can explicitly apply the validated saved CEM draft or select an authored preset. None of these actions publishes to the realtime voice. Automated Gauntlet cases cover budget, terrain, corrupt inputs, unchanged ranking, source provenance, basket consent and disabled live transactions. Human evaluation of soul and conversational quality remains required.
/demos/nike/runs reviews a fictional easy 5 km and five fictional 1,000 m repetitions. Times, range and target deviations are calculated. Pre’s interpretation is authored, distinguishes observation from inference, and asks for effort, recovery, conditions and session purpose. It does not infer threshold, VO₂ max, clinical risk or race readiness from pace alone. These examples show desired depth for novice and experienced runners; they are not validated professional coaching.
Nike+ / NRC data access
Research on 5 October 2026 did not establish a currently supported public Nike+ developer API. Direct member/run-data access remains a Nike-authorised integration requirement. Nike’s Strava partnership provides a supported way for members to share NRC workouts to Strava; it does not give this prototype Nike API access.
Strava’s API Policy, effective 1 June 2026, restricts ordinary API data from AI operation, including context ingestion and derived data. Its official MCP exception permits personal use of one’s own data, while excluding commercial or third-party access outside personal use. This is not a blanket permission to feed NRC-through-Strava data into commercial Pre. An approved agreement and supported integration are needed. Sample data is the present demonstration route; user-entered summaries and original device exports would need their own consent, provenance and processing design before integration. No scraping or credential workaround has been added.
Shoe photographs and personal Strava access — 5 October 2026
The shoe encounter now uses actual Nike product photography for Winflo 11, Pegasus 41 and Pegasus Trail 5, both in recommendation cards and the selected practice basket. Local JPEG copies keep this prototype reliable. Each catalogue entry retains the exact Nike CDN image URL in imageSource; imageAlt describes the pictured colourway. Photography remains Nike’s material; this illustrative prototype does not establish rights for a public commercial launch. Images do not imply current stock, an available colourway, price or fit. The Pegasus 41 photograph was sourced from https://www.nike.com/id/t/pegasus-41-mens-road-running-shoes-RZm89S/FD2722-102; the other photographs came from the product pages already recorded in the catalogue. These referenced colourways were marked unavailable when checked.
The run page now explains the personal-data ambition and links the official route, without a fake connection button or collecting credentials. Strava’s official subscription MCP is read-only, scoped to the individual and revocable; its help page currently documents Claude, not this Pre website: https://support.strava.com/en-us/articles/15401531-what-is-the-strava-mcp-connector. The June 2026 API policy sections 3.5 and 5.3 distinguish the personal MCP exception from commercial third-party access and ordinary API data used as AI context: https://www.strava.com/legal/api_policy. Personal consent or naming data “personal” does not establish permission for our commercial integration. Next step: obtain written permission covering this use case, then design scoped OAuth, revocation and minimal per-user data access. No Strava account connection has been implemented.
Specialist depth beneath the personality
The whitepaper’s six facets of an expert persona record the proposed specialist architecture: strong reasoning, a searchable expert library, personal memory with permission, reliable analysis tools, natural voice and expert evaluation. Pre’s priority is running and competition, including the mental aspects; Arthur’s is rugby. These facets extend the persona’s usefulness without replacing its character or widening its permissions. Retrieval, persistent memory and athlete-data integrations remain proposed, not connected capabilities.