The Trio Model
Musketeer separates work into three cognitive roles: Originator, Executor, and Cross-Examiner. This is not three tools. It is three kinds of thinking, each suited to a different model.
The triangle of responsibility
Originator
/ \
/ \
/ intent \
/ + context \
/ \
Cross-Examiner ----- Executor
observation output Each vertex of the triangle has a distinct cognitive responsibility. Information flows between them through structured handoffs. The human sits at the center, deciding when and what to hand off.
Originator
The Originator is conversational. It holds the long-running context of what you are trying to accomplish. It is ambiguity-tolerant. It helps you think through the problem, form intent, and articulate constraints.
The Originator is typically a chat-style model. ChatGPT excels here. Claude in conversation mode excels here. These models are built for dialogue, for exploration, for refining ideas through iteration.
What the Originator does:
- Holds the ongoing conversation about what you want
- Helps you articulate intent and constraints
- Refines ideas through dialogue
- Prepares handoffs for the Executor
What the Originator does not do:
- Execute bounded tasks
- Generate production artifacts
- Verify its own suggestions
Executor
The Executor is bounded. It receives a clear task, executes it, and produces output. It is instruction-following. It does not ideate. It does not explore alternatives. It does what it is told, within the constraints it is given.
The Executor is typically an agentic or code-focused model. Claude Code excels here. Codex excels here. These models are built for bounded execution, for following instructions precisely, for producing artifacts.
What the Executor does:
- Receives structured handoffs with clear intent
- Executes bounded tasks
- Produces artifacts (code, text, files)
- Reports what it did
What the Executor does not do:
- Question the overall strategy
- Explore alternative approaches
- Hold long-running context
- Verify its own output
Cross-Examiner
The Cross-Examiner is observational. It looks at what the Executor produced and reports what it sees. It is detached. It does not suggest fixes. It does not write code. It observes and reports.
The Cross-Examiner is typically a model with strong analytical capabilities. Gemini excels here. Any model can serve this role when prompted correctly. The key is that it has no stake in the output.
What the Cross-Examiner does:
- Reviews Executor output against stated intent
- Identifies discrepancies and concerns
- Reports observations without suggestion
- Provides an independent perspective
What the Cross-Examiner does not do:
- Fix problems it finds
- Write code
- Participate in the creative process
- Make execution decisions
Why not two roles? Why not four?
Two roles collapse the distinction between observation and execution. When the same model that executes also verifies, it has a stake in confirming its own work. The verification becomes performative.
Four or more roles add complexity without adding cognitive value. The trio maps to fundamental cognitive acts: forming intent, executing intent, and checking execution. Adding more roles tends to blur rather than clarify.
The trio is minimal and sufficient.
Not all roles are always used
For simple tasks, you may only use Originator and Executor. For exploratory work, you may only use Originator. The Cross-Examiner is most valuable when the stakes are high or when you have reason to distrust your own judgment.
The trio is a framework, not a requirement. Use what you need.
The separation is cognitive, not technical
You could run all three roles on the same model. The results would be worse, but it would work. The separation is about cognitive responsibility, not technical architecture.
When you use different models for different roles, you get the benefits of specialization. But the core insight is about separation, not about which specific model fills which role.