LAST PROMPT ENGINE — TECHNICAL OVERVIEW

The Engine

Last Prompt is not just a game. It is a decision-intelligence evaluation engine that can be wrapped in any thematic skin — survival, corporate, diplomatic, scientific.

The engine evaluates the quality of human reasoning under uncertainty. It is content-agnostic, skin-agnostic, and domain-agnostic. The only constant is the rubric.

Engine Thesis

The scarce skill is no longer information recall.

AI-mediated systems are increasing the complexity of human decision-making at every level. The bottleneck is not access to information — it's the ability to reason structurally under uncertainty.

Last Prompt is built on a single thesis: better reasoning produces better outcomes. The engine proves this by making the quality of your written plan the direct cause of what happens next in the simulation.

The AI does not drive, decide, or progress the simulation. It only evaluates how well the player thought through the problem.

ENGINE_CORE — EVALUATION_FLOW.pseudo

function evaluatePlan(plan, state) {

// Score against 5-criterion rubric

scores = rubric.evaluate(plan, state);

band = getQualityBand(scores.total);

// Apply interdependency multipliers

deltas = calculateDeltas(band, state);

// Set flags for future crises

flags.update(scores, plan);

return { scores, band, deltas, narrative };

}

// The engine never hardcodes variable names.

// All labels are pulled from the active skin config.

Cognitive Complexity Scaling

Variable count is not cosmetic.

The number of active variables in a skin directly determines the cognitive complexity of the simulation. The engine supports any number.

2–3 Variables

Ethical Compression

Binary trade-offs, moral tension. Fewer variables amplify the emotional weight of each decision.

e.g., Security vs. Compassion

4–6 Variables

Systems Leadership

Interdependency and prioritisation. Decisions ripple across multiple systems simultaneously.

e.g., Colony Survival, Corporate Crisis

7+ Variables

Executive Strategy

High-complexity environments requiring abstraction, delegation, and long-horizon thinking.

e.g., National Crisis, Diplomatic Simulation

Architecture

Engine vs. Skin

The engine is the unseen hand. The skin is the sensory experience. They are completely decoupled.

LAYER 01

The Engine (System Logic)

Content-agnostic. Never uses the words "Food", "Health", or "Colony". Pulls all labels from the active skin config.

Stat Handler

Manages a dynamic list of variables (Stats) with min/max clamping and threshold logic. The engine never hardcodes variable names — it reads them from the active skin.

Evaluation Orchestrator

Manages the interface with the AI backend and the 0–10 rubric scoring system. Runs at temperature 0 to ensure deterministic, consistent evaluation.

Rule Engine

A deterministic loop that filters crisis events based on stat thresholds and flags set by previous decisions. Your history shapes what comes next.

Decision Loop

Input → AI Analysis → Outcome Resolution → Next Crisis Selection. The same loop runs in every skin, every cycle.

Memory & Flags

Decisions set hidden flags that persist across cycles. A strong decision in Week 1 can unlock opportunities in Week 3. A poor one can trigger cascading crises.

LAYER 02

The Skin (Thematic Content)

The sensory experience and context. Defined entirely in JSON — swappable without touching engine code.

Thematic Vocabulary

Defines whether the simulation is Colony Survival, Corporate Strategy, or Diplomatic Crisis.

Data Collections

The specific events.json, deltas.json, and narratives.json that populate the world.

Character Profiles

Each advisor has an archetype, core fear, hidden doubt, generational lens, and decision bias weights.

Visual Styles

CSS variables (colours, fonts, layout) that represent the world's atmosphere.

Stat Mapping

Maps generic engine keys (Stat_01, Stat_02) to human-readable labels for the skin's context.

Data-Driven Variable Mapping

ENGINE KEY	COLONY SKIN	CORPORATE SKIN
Stat_01	Sustenance	Cash Flow
Stat_02	Health	Employee Well-Being
Stat_03	Security	Regulatory Compliance
Stat_04	Cohesion	Team Engagement
Stat_05	Infrastructure	Operational Infrastructure
Time_Unit	Week	Quarter
Entity_Name	The Colony	The Enterprise

AI Guardrails

The evaluator cannot be gamed.

Substantial guardrails prevent players from gaming the system, asking for full marks, or exploiting the AI's tendency to be agreeable.

Temperature: 0.0

The evaluator runs at zero temperature. No creative drift. The same plan gets the same score every time.

No Pity Points

The AI must not assume positive outcomes unless the player explicitly describes the mechanism. Vague plans are penalised.

Harsh Interpretation

Plans under 20 words, or lacking contingencies, are immediately penalised. The evaluator is not a cheerleader.

Reasoning Required

Every rubric score must include a reasoning string. The evaluator is accountable for every point it awards or withholds.

Metric Masking

Advisors never reference numeric outcomes. They think in human consequences: 'Morale will shatter' — not '+2 Cohesion'.

Domain Containment

Specialists only see the world through their role. A Security advisor cannot comment on social cohesion. Advice is humanly incomplete by design.

Collaborators

Have a domain? Build a skin.

The engine is modular. If you work in medicine, diplomacy, urban planning, education, or any field where structured reasoning under uncertainty matters — the Last Prompt Engine can be adapted to your context.

We're looking for collaborators who are frustrated by polarised thinking and inspired by the idea of lateral reasoning as a trainable skill.

Define your variables (2–10+)

Write your crisis events and character profiles

The engine handles evaluation, scoring, and consequence resolution

See the engine in action.

Explore the two live skins currently in development.

Colony Corporate