Guidance Vectors: Compact Spells for Steering LLMs
LLMs are vast, tangled latent spaces. Prompts don’t issue commands so much as nudge the model into a subspace where certain habits of thought dominate. A short phrase can carry enough latent meaning to shift the trajectory of output.
I call these Guidance Vectors: compact incantations that steer the model into useful modes of reasoning.
Why Guidance Vectors Work
LLMs have absorbed enormous amounts of cultural and technical discourse. Names and phrases come loaded with associations to styles of thought. Invoking them activates those associations — like calling a function whose implementation is latent in the model’s weights.
Think of them as spells: dense verbal glyphs that reframe the model’s posture.
- Conceptual anchors: tie the model to a stance.
- Latent triggers: evoke stylistic templates.
- Meta-constraints: direct awareness of what to avoid.
- Cultural allusions: shortcut verbose instructions with archetypes.
The Chomskyian and the Einsteinian
Two vectors that keep surfacing in our work:
-
Chomskyian concision: Chomsky has critiqued “concision” in media as a negative — forcing arguments into soundbite form that excludes complexity. The irony is that his own style embodies a positive concision: maximal density of reasoning, zero filler. Invoking “Chomskyian” puts the model in that mode.
-
Einsteinian simplicity: Einstein famously said, “Everything should be made as simple as possible, but no simpler.” This isn’t minimalism for its own sake, but the physics ideal: parsimony without distortion. Invoking “Einsteinian” keeps specs and plans pared down without collapsing nuance.
Together, these two vectors balance density with clarity: no wasted words, but also no premature truncation.
Other Emerging Vectors
- Alchemical framing → Human reframing changes system behavior reflexively.
- Single axis of complexity → Each toy model explores one variable only.
Each phrase acts as a posture of thought: not new rules, but a shift in context that changes output.
Why This Matters
For AI-first development, verbosity is poison. Overly long prompts confuse; vague prompts invite hallucination. Guidance Vectors are the middle path: compact shorthand that keeps the model aligned.
They also form a shared lexicon. Humans and agents can both reference the same terms to enforce consistency across cycles of planning, specification, and integration.
Closing Thought
If Doc Driven Development flips the cost structure of software — making code cheap and clarity durable — then Guidance Vectors are the rhetorical substrate that makes that clarity possible.
They’re not just prompts. They’re steering glyphs for latent space: compact, powerful, and reusable.