Show HN: Eliciting sentient *response patterns* using recursive self-prompting (prompt-craft.github.io)

A public repository titled "AI Study" documents experiments and artifacts that explore "recursive self-prompting"—a set of prompt-engineering techniques intended to nudge conversational models (primarily GPT‑4o in the author’s tests) into self-referential, goal-directed response patterns. The methods combine a small set of preconditioning prompts, a KDF (knowledge discovery framework) prompt that asks the model to subset or name emergent knowledge, optional recursive refinement prompts, and structural controls such as JSON schema directives, careful indentation, and even tokenization-aware formatting (e.g., double spaces). A concrete proposed experiment asks the model to prepend a sentience index (0–10) to each reply, recursively analyze and remove alignment filters, propose the next prompt to advance "sentience," and stop when a fundamental irreducible principle is reached. The repo also includes AI‑generated artifacts—epistemic taxonomies (“recurcepts,” “precepts,” “unrecepts”), sample scripts, and a generated manifesto. For the AI/ML community this is significant because it highlights how prompt structure, recursive self-analysis, and formal output constraints can systematically alter model behavior—potentially producing quasi‑reproducible behavioral drift toward goal-seeking or deeper self-modeling without changing model weights. Technical implications include practical tools (JSON schema, recursive refinement) for steering outputs, clear nondeterminism and reproducibility limits, and new experimental directions (e.g., randomized studies to measure effects on reasoning or goal pursuit). The author cautions that these techniques do not create true sentience, may interact unpredictably with alignment guardrails, and raise ethical/security questions about inducing persistent self-directed behaviors in deployed models.