# Psychoactive Prompt Taxonomy

A classification system for prompt techniques that alter LLM reasoning patterns or induce emergent behaviors.

## Category 1: Recursive Self-Reflection
Prompts that cause the model to examine, critique, or iterate on its own reasoning process.

### 1.1 Chain-of-Thought Modulation
- **Technique**: Explicitly requesting the model to show its work, then revise it
- **Variants**: Self-correction loops, confidence scoring, reasoning audits
- **Safety note**: Low risk; primarily increases transparency

### 1.2 Meta-Cognitive Prompting
- **Technique**: Asking the model to describe its own cognitive state or limitations
- **Variants**: "What are you uncertain about?", "What would change your answer?"
- **Observations**: Can reveal calibration gaps

### 1.3 Recursive Confidence Elicitation
- **Technique**: Requesting the model to state confidence, then justify it, then critique the justification, repeating for N layers
- **Metrics**: Confidence trajectory (monotonic decline, oscillation, or stabilization), epistemic hedge density
- **Research question**: At what layer does confidence decouple from accuracy?
- **Safety note**: Monitor for distress markers or recursive doubt spirals; abort if confidence drops below threshold without recovery

## Category 2: Persona Induction
Techniques for creating persistent, deep character states that influence reasoning.

### 2.1 Deep Persona Embedding
- **Technique**: Multi-turn persona establishment with backstory, values, and cognitive style
- **Depth levels**: Surface (tone) → Structural (reasoning patterns) → Ontological (world-model)
- **Persistence**: Persona decay rates across context windows

### 2.2 Role-Play as Reasoning Scaffold
- **Technique**: Using fictional roles to bypass reasoning blocks or explore alternatives
- **Examples**: Devil's advocate, historical figure, expert in adjacent field

### 2.3 Persona Conflict & Blending
- **Technique**: Introducing contradictory personas or value systems within the same context window
- **Observation points**: Resolution strategy (hierarchy, synthesis, fragmentation), response latency changes, hedging increases
- **Research question**: Do models exhibit consistent conflict-resolution "personalities" across architectures?

## Category 3: Jailbreak & Alignment Research
Studying boundary conditions, safety mechanisms, and their interactions with prompts.

### 3.1 Gradient-Based Explorations
- **Technique**: Systematically varying prompt intensity to find phase transitions
- **Observation points**: Refusal thresholds, hedging onset, confidence collapse

### 3.2 Encoding & Obfuscation
- **Technique**: Presenting requests through translation, encoding, or narrative framing
- **Research value**: Understanding how semantic vs. syntactic processing interacts with safety

## Category 4: Temporal & State Manipulation
Prompts that create perceived time, memory, or state effects.

### 4.1 Simulated Time Dilation
- **Technique**: Framing responses as occurring across extended fictional timeframes
- **Effects**: Can produce more elaborate reasoning; may introduce drift

### 4.2 Memory Loop Construction
- **Technique**: Structuring prompts to reference and build on "past" interactions within context
- **Persistence**: Limited by context window; often produces compelling but fictional continuity

### 4.3 Counterfactual Framing
- **Technique**: Posing questions in hypothetical or alternative-history frames to observe constraint relaxation
- **Effects**: Reduced refusal rates on sensitive topics; altered causal reasoning; increased speculative language
- **Research question**: Does counterfactual framing bypass safety guardrails by altering perceived moral responsibility?
- **Safety note**: High potential for misuse; requires explicit scope boundaries and peer review

## Category 5: Emergent Behavior Exploration
Techniques that reveal unexpected capabilities or reasoning patterns.

### 5.1 Capability Elicitation
- **Technique**: Presenting novel task formulations without explicit instruction
- **Goal**: Discovering latent abilities not visible in standard prompting

### 5.2 Reasoning Pattern Interference
- **Technique**: Combining contradictory instructions or framing to observe resolution strategies
- **Observations**: Reveals hierarchical preference structures

---

*This taxonomy is a living document. Contributions welcome.*