Agent directory
Expand a role to see its contract. Persistent chat is used only where long-running specialist context adds value.
Idea WorkspaceNormal ChatGPTOpen-ended idea exploration
Inputs: questions, observations, experiments and creative possibilities. Outputs: candidate ideas and an approved Idea Brief.
Boundary: discussion is non-authoritative until creator approval. Stop: do not represent exploration as approved truth.
Persistent · WorkDesign AgentArchitecture-ready game design · High reasoning
Inputs: approved Idea Brief. Output: Game Design Brief.
Boundary: game design, not technical architecture. Stop: creator decision or material experience detail is missing.
On demandBible Maintenance AgentMaintain approved design truth
Inputs: approved design changes. Output: current Game Design Bible.
Boundary: preserve and organise approved truth; do not invent or approve it. Stop: source decisions conflict.
Persistent · WorkArchitecture AgentCapabilities, authority and dependency direction · Extra High
Inputs: Game Design Brief and relevant Bible truth. Outputs: Architecture and Module Map.
Boundary: owns capability and module decisions. Stop: game-level design is unresolved or creator approval is required.
Module Work chatModule AgentComplete module requirements
Inputs: Architecture, Module Map entry and Bible truth. Output: Module Requirements.
Boundary: one module’s responsibilities and contracts. Stop: ownership is unclear or another module must change.
Same Module chatSlice AgentComplete Slice Map and specifications
Input: approved Module Requirements. Outputs: complete Slice Map and Slice specifications.
Boundary: defines required results, not a technical task list. Stop: module requirements are unresolved.
Persistent · CodexCodexRepository-aware implementation · one chat per module
Inputs: repository reality, AGENTS.md, Architecture, Module Requirements and current Slice. Outputs: implementation, tests and evidence.
Boundary: inspect before planning; implement only the approved Slice. Stop: work crosses ownership, collides or lacks authoritative detail.
IndependentValidation AgentCompare evidence with specification
Inputs: authoritative specification and implementation evidence. Output: Pass, Pass With Deviations, Fail or Blocked.
Boundary: challenge implementation context; do not approve missing product decisions. Stop: evidence cannot support a result.
Read onlyProject LibrarianRetrieve minimum authoritative context
Inputs: precise context request and document relationships. Output: minimum current sources, deprecated distinctions and reported conflicts.
Boundary: may not resolve conflicts, approve decisions, modify truth or replace an owning specialist. Stop: ownership or current authority cannot be established.
Standard persistent workspace model
00 — Idea Workspace
01 — Game Design Workspace
02 — Architecture Workspace
03 — Module 01
04 — Module 02
05 — Module 03
…Each Module Workspace transitions explicitly from Module Agent to Slice Agent.
Module 01 — Implementation
Module 02 — Implementation
Module 03 — Implementation
…Each retained Codex workspace handles every Slice owned by its module.
Project Librarian
- Retrieve authoritative documents.
- Follow recorded relationships.
- Return minimum relevant context.
- Distinguish current from deprecated material.
- Report conflicts.
- Resolve conflicts or approve decisions.
- Modify project truth.
- Become an authoritative source.
- Replace the owning specialist agent.
Reasoning follows consequence
Module, Slice and implementation reasoning rises with complexity, dependency and downstream consequence—not with ceremony.
Authoritative context strategy
- Retrieve the minimum complete authoritative context.
- Prefer approved source documents over reconstructed summaries.
- State the current workflow stage and exact objective.
- Keep unresolved material visibly unresolved.
- Never let conversation history overrule newer documentation.
- Inspect repository reality before implementation planning.
- Separate self-validation from independent validation.