Radical Geek field note
The Hypnotist's Prompt: What 100 Years of Trance Science Tells Us About Engineering AI
Prompt engineering and hypnosis both shape complex systems through perspective, context, and permission. That parallel has practical lessons for AI engineering.
Originally published on LinkedIn on 20 March 2026.
Two fields arrived at the same destination by different roads.
Hypnosis has spent more than a century studying how language, framing, permission, and context shape the behaviour of a complex system without directly forcing it. Prompt engineering discovered a similar problem from the other side: how do you shape a language model’s internal state so it produces a useful output?
The mechanisms are different. Humans are not language models. Language models are not conscious minds. But the functional parallels are too useful to ignore.
Perspective, Context, Permission
In conversational hypnosis there is a useful triad: perspective, context, permission.
Perspective is the lens through which a subject interprets the situation. Context is the frame that determines what meanings are available. Permission is the internal gate that decides whether a suggestion can be accepted.
AI systems have equivalents.
The system prompt installs a perspective. The context window establishes the frame. The model’s alignment, training, and immediate instructions influence what kinds of outputs feel available or unavailable.
You are not commanding the system by force. You are shaping the world it thinks it is operating inside.
Resistance Is Information
When a model resists a request, engineers often treat that resistance as an obstacle. The hypnosis parallel suggests another view: resistance tells you the frame the system is using.
If the model says it cannot do something because of safety, ambiguity, missing authority, or lack of context, that response is diagnostic. It tells you what condition has not been satisfied.
The better move is not to fight the resistance. It is to reframe the task truthfully so the model can see why the request is appropriate, bounded, and useful.
Pacing Before Leading
Chain-of-thought and step-by-step prompting work partly because they pace before they lead.
You start with premises the model can accept. Then you build a path. By the time the model reaches the conclusion, it is not jumping; it is completing a trajectory.
This is not politeness. It is mechanics.
The same idea appears in good architecture work. You do not ask a team to accept a conclusion before you have walked through the constraints, trade-offs, and evidence. Models respond to that structure too.
Prompt Injection Is the Shadow
The uncomfortable part is that prompt injection uses the same levers adversarially.
A malicious instruction tries to install a new perspective, create a hostile context, and grant permission for behaviour the system should refuse. It is not magic. It is context displacement.
That suggests a defensive direction: do not rely only on brittle rules. Build stronger identity, clearer source boundaries, and a more stable frame for the model. Make it harder for untrusted content to redefine the system’s role.
The Useful Lesson
The point is not that AI is hypnosis. The point is that both domains have learned something about complex generative systems: direct instruction is weaker than well-designed context.
If you want better AI output, do not just polish the prompt. Shape the perspective. Curate the context. Make permissions explicit. Reduce ambiguity. Protect the frame.
That is where prompt engineering grows up into context engineering.