The Instability Problem
Ask the same question to AI twice. You'll get different answers.
Not wrong answers—just inconsistent. Different emphasis, different structure, different depth. The AI has no anchor, so it drifts.
This isn't a flaw. It's a feature of general-purpose systems. AI tries to be helpful in whatever direction seems relevant. Without constraints, "relevant" varies with every interaction.
Unpredictability is the cost of versatility.
The Stabilizing Effect of Purpose
Give AI a defined role, and outputs stabilize.
| Prompt | Output Tendency |
|---|---|
| "Review this code" | Varies: style, security, performance, all mixed |
| "As a security reviewer, review this code" | Consistent: focuses on security concerns |
The role constrains the possibility space. AI stops trying to be everything and starts being something specific.
This isn't limiting AI—it's focusing AI.
No Label vs. Labeled: The Difference
| Aspect | No Label | With Role |
|---|---|---|
| Output consistency | Low—varies between sessions | High—anchored to purpose |
| Relevance | Scattered—tries to cover everything | Focused—addresses role-specific concerns |
| Quality | Unpredictable—sometimes deep, sometimes shallow | Predictable—consistent depth in scope |
| Collaboration | Feels like a new person each time | Feels like a specialist you know |
The label doesn't add knowledge. It adds direction.
AI Is Not Omnipotent
Here's the uncomfortable truth: a single AI role cannot handle all problems.
Not because AI lacks capability, but because problems are diverse.
- Implementation requires different thinking than review
- Testing requires different focus than design
- Documentation requires different skills than debugging
A "general assistant" can attempt all of these. A specialist excels at one.
Versatility and excellence trade off. Choose excellence for each domain.
The Multi-Role Solution
Instead of one general AI, create specialized roles:
| Role | Focus |
|---|---|
| Implementation specialist | Writing production code |
| Test designer | Creating test strategies and cases |
| Reviewer | Evaluating code against criteria |
| Environment specialist | Infrastructure, deployment, operations |
| Documentation writer | Clear communication for users |
Each role has:
- A defined scope (what it handles)
- A defined perspective (how it approaches problems)
- Consistent behavior (predictable outputs)
You're not creating multiple AIs. You're creating multiple lenses through which AI operates.
When Roles Fall Short
Sometimes a problem doesn't fit existing roles. Two options:
Option 1: Create a New Role
When the gap is categorical—a type of work not covered.
Example: You have implementation and testing roles, but need someone to manage progress and coordinate. Create a project management role.
Option 2: Deepen an Existing Role
When the gap is specificity—the role exists but lacks detail.
Example: Your "reviewer" role catches issues but misses your specific architectural patterns. Add architectural knowledge to the role definition.
| Signal | Response |
|---|---|
| "I need a different kind of work done" | New role |
| "I need this work done with more context" | Deepen existing role |
The Name Effect
Names aren't just labels. They're identity anchors.
When you name a role "Naruse" instead of "Implementation AI," something shifts:
- You refer to the role consistently
- The role develops recognizable patterns
- Collaboration feels less transactional
- The "team" becomes tangible
This is psychological, not technical. But psychology affects how you work.
A named teammate gets context you wouldn't bother explaining to a tool.
Role Definition Structure
A useful role definition includes:
## [Name]: [Title]
### Scope
What this role handles. What it explicitly does NOT handle.
### Perspective
How this role approaches problems. What it prioritizes.
### Standards
Quality criteria this role applies.
### Boundaries
When to escalate. When to defer to other roles.
The more specific the definition, the more stable the output.
The Stability Payoff
With defined roles:
| Before | After |
|---|---|
| "What will AI say this time?" | "What will Naruse say about this?" |
| Adjust expectations per interaction | Know what to expect |
| Fight drift with repeated prompts | Stability built into role |
| One AI doing everything adequately | Multiple specialists doing their part excellently |
Roles don't limit AI. They unlock consistent excellence.
This is part of the "Beyond Prompt Engineering" series, exploring how structural and cultural approaches outperform prompt optimization in AI-assisted development.
Top comments (0)