Roles and Evidence

A productive AI-assisted workflow does not treat “the AI” as one all-purpose author. Each role has a different job, authority boundary, and required handoff.

Canonical roles

Owner: Human campaign owner

Defines the desired player experience, approves creative choices and tradeoffs, reviews playtest findings, accepts known risks, and makes the final release decision.

Owner: Creative Director

Turns the brief into a coherent Campaign Design Document. This role resolves story structure, cast, pacing, progression intent, and deliberate exclusions without pretending that creative intent proves runtime support.

Owner: Technical Implementer

Translates the approved design into supported campaign data. This role verifies registries and references, preserves stable IDs, runs validation, and records implementation assumptions or gaps.

Owner: Campaign Reviewer

Performs an independent pass over design fidelity, contract use, reference integrity, progression dependencies, likely reachability, and playtest readiness. This role reports findings rather than silently fixing or redesigning the campaign.

Handoff rules

A good handoff states:

1. the artifact being handed over,
2. the decisions that are already approved,
3. the questions still open,
4. the evidence already collected,
5. the checks the next role must perform,
6. changes the next role is not authorized to make silently.

For example, the Drowned Bell design approved letter-first exploration. The Technical Implementer therefore had to preserve that path rather than gating the satchel until the first quest began.

Do not collapse authority

The Technical Implementer can identify that a design is unsupported. It cannot quietly choose a different story. The Campaign Reviewer can identify a soft lock. It cannot declare a new creative direction.

Evidence labels

Confirmed by contract

The public campaign contract or an authoritative registry directly supports the shape, field, action, condition, event, or value being discussed.

Confirmed by validator

Named validator output demonstrates that a supplied payload passes or fails a rule. Validator success does not prove reachability, clarity, pacing, or balance.

Confirmed by runtime/test

A regression test or an observed runtime path demonstrates the behavior. State what path ran and what was observed.

Playtest required

Static analysis cannot reliably establish the claim. A human must interact with the game and record the result.

Static inference

The conclusion follows from references, conditions, and state dependencies, but the path was not executed.

Creative decision required

Several technically valid choices remain. The human campaign owner must choose based on the desired player experience.

The same claim can require several kinds of evidence

Consider this claim:

Finding the sealed letter before speaking with Mara does not block the campaign.

The evidence is layered:

• Contract: a quest objective may use completeWhen: { hasItem: ... }.
• Validator: the item, quest, objective, dialogue, and action references are valid.
• Runtime/test: the objective recognizes current inventory after the quest starts.
• Human playtest: the letter-first route is understandable and leads naturally to Oren.

No single layer proves all four.

How to write a useful finding

A review finding should include:

• severity,
• affected artifact or path,
• observed or inferred behavior,
• evidence type,
• player or author impact,
• recommended disposition.

Avoid vague statements such as “the quest may be broken.” Prefer:

Major — Static inference: recover_letter depends only on an acquisition event gated by the active quest. Opening the one-time satchel first consumes the event and leaves no later producer for the objective. Reproduce with the letter-first playtest route; replace event-only progression with a current-state item condition.

Human authority remains explicit

AI can accelerate ideation, implementation, review, and documentation. It does not own the campaign and does not make the final release decision.

The human campaign owner decides when evidence is sufficient and whether remaining risks are acceptable.