Manifold Prompting (MP)

Manifold Prompting (MP) is a user-side, multi-turn method for sustaining stable, low-friction collaboration with LLMs across long-horizon work

What it is:

Manifold Prompting is repeatable interaction method for keeping constraints, standards, and working assumptions alive across many turns, even as the task expands, compresses, and changes shape.

Where prompt engineering optimises a prompt, MP optimises the interaction over time: how context is introduced, how it is tightened, how constraints are enforced, and how continuity is maintained without constant re-anchoring.

Understanding the concept:

In long-horizon sessions, the failure mode is rarely “the first answer.” More often, it’s continuity decay: tone flattens, constraints quietly stop carrying forward, the model drifts into generic defaults, and the user ends up repeating the same standards again and again.

Manifold Prompting addresses that stability problem by treating the conversation as an accumulative system rather than a sequence of isolated completions. The method works by bringing three user-side stabilisation functions together (the Stability-Enabling Triad):

  • Stable Affective Manifold (SAM): a register the user can sustain and the model can reliably mirror without misrouting.

  • Multi-Layer Coherence (MLC): alignment between what the user is doing semantically, what they are signalling about the task, and the affective register they’re using.

  • Structured Turn Geometry (STG): predictable update rhythm across turns (expansion, compression, stabilisation), plus turn-linking moves that keep the interaction chained rather than reset-prone.

Operationally, MP shapes a continuation corridor: the narrowed subset of plausible next responses that remain consistent with the user’s accumulated signals and constraints. Over time, this reduces interpretive variance, improves constraint carry-forward, increases compression tolerance (saying less without losing precision), and strengthens drift resilience.

MP is structurally aligned with High-Coherence Interaction State (HCIS): HCIS names the emergent regime observed in high-performing long-horizon dyads; MP specifies a user-expressible procedure that can increase the likelihood of converging toward, and sustaining, HCIS-consistent continuation.

Website version of the preprint: https://www.annawojewodzka.com/manifold-prompting-full-paper
DOI preprint here: 10.5281/zenodo.18777254
(This paper is an interaction-level method description and evaluation proposal, not a model architecture claim or a controlled user study.)

ORCID: https://orcid.org/0009-0001-9458-7150