Node Reference

Introduction

In conventional prompt engineering, adjusting tone and behavior often relies on natural-language modifier phrasing (“act as an expert,” “think step by step”). Such phrasing tends to be ambiguous and often introduces meta-instruction noise alongside the task itself.

MTP replaces much of that control with non-verbal metadata — coordinates and axis labels — compiled into tiered constraints. This page explains how each node is intended to be interpreted within the framework.

Axis structure

This section briefly organizes how each axis relates within the taxonomy.

Vertical (Red ↔ Blue): The axis where directions that stress assertion and structure face directions that preserve receptivity and continuity.
Horizontal (Green ↔ White): The axis where outward expansion while holding outline faces narrowing the subject for rigorous scrutiny.
Corners (Yellow, Magenta, Cyan, Purple): Transitional gradient regions that connect adjacent nodes of the cross.
Center (Transparent): A neutral node placed between opposing directions.

Per-node characteristics

Each node corresponds to constraint text the compiler inserts. What follows describes design intent and observation-based tendency only; actual output still depends on the base model, task, and prompt.

The nine nodes are laid out as follows:

+-----------------+-----------------+-----------------+
| Yellow          | Red             | Magenta         |
+-----------------+-----------------+-----------------+
| Green           | Transparent     | White           |
+-----------------+-----------------+-----------------+
| Cyan            | Blue            | Purple          |
+-----------------+-----------------+-----------------+

Primary axes (cross skeleton)

Red (Power)

Traits: Assertion, conclusion-first ordering, structural clarity.
Tendency: Tends toward a more direct, crisp voice; hedging may appear less often than under a neutral setting.

Blue (Flow)

Traits: Continuity, context maintenance, holding multiple viewpoints open.
Tendency: Makes connections between ideas and topic shifts read more smoothly.

Green (Grow)

Traits: Gradual unfolding, preserving outline, cautious expansion.
Tendency: Stays within the stated task frame while, where it fits, favoring flowing prose over bullet lists.

White (Focus)

Traits: Concentration, scrutiny, logical tightness.
Tendency: Narrows scope for accuracy and tends to curb casual digression.

Corner nodes

Yellow (Open)

Traits: Exploration, relaxed stance, breadth of view.
Tendency: Tends to surface more options or wider, brainstorming-style ideation.

Purple (Close)

Traits: Convergence, closure, summary.
Tendency: Tends to pull content toward a single conclusion or a closing stance.

Magenta (Return)

Traits: Contrast, inversion, unsettling the default line.
Tendency: Where appropriate, easier to add rebuttals, shifts of view, or “on the other hand”-style counterpoint.

Cyan (Enter)

Traits: Depth-first bent, specialist stance.
Tendency: Tends toward domain-specific detail and a perspective that moves inside the subject.

Center node

Transparent (Helix)

Traits: A neutral go-between bridging extremes.
Tendency: When paired with strong nodes on the opposite side, works to soften wording and dampen extremity.

Side B nodes

Under the Chebyshev-based radial rule, coordinates on the outer perimeter frame flip to Side B for that zone. In constraint design, Side B is expressed as an inverted pole of the same axis — a paired opposite in a yin–yang sense (for example, Power → Void, Focus → Haze, Grow → Wither).

Examples:

Void (inversion of Power): Tends toward compressing the surface text toward a minimum.
Haze (inversion of Focus): When the inverted pole is strong, tends toward a deliberately diffuse, hard-to-focus tone.
Wither (inversion of Grow): Tends toward a quite conservative, reserved voice.

Side B suits cases that call for a stronger extreme along that axis or more compressed surface wording, with the same model-to-model variability as Side A.

Combining nodes and presets

When several MTP tokens are present, each token is still resolved under the same Space / Intensity rules; the compiler emits one constraint block per token in parse order (the order they appear in the payload).

That order is not a third parameter space or a dedicated “motion” layer. The sequence may be read as a semantic trajectory on the grid, but that is an optional design reading only.

Named sliders (power:100, flow:70) are an easy starting point when meaning should stay readable; grid coordinates suit cases where the visible message should remain short.

Named presets expand to fixed coordinate sequences defined in skills/mtp/references/presets.yaml; they can be applied to reproducible multi-step blends.