Persona is the stable who
The AI Agent persona carries the analytical identity, cognitive style, archetype, subject heads, and model/backend assignment. It should not change every time a task changes.
Prompt assembly
How iPulse turns reusable methodology components into forecast prompts
A forecast prompt is not handwritten from scratch. iPulse assembles it from versioned components so each prediction can be traced back to the agent, mode, task, input data, output schema, and context used to produce it.
Updated July 3, 2026
Overview
Prompt assembly is a controlled build process
iPulse uses task configurations to assemble prompts from reusable components. This makes the system easier to audit than a one-off prompt. The same AI Agent can be paired with different modes, tones, registers, data inputs, and output formats while preserving a traceable record of what generated the forecast.
The docs explain the public structure and purpose of prompt assembly. They do not publish the full production prompts, private component text, or operational secrets.
What improved
Prompt Architecture v2 separates identity, execution, and task
A key lesson from operating AI forecasts is that too many concepts can accidentally collapse into one prompt. If persona, model, mode, horizon, output schema, tone, and subject instructions all live as one hand-edited blob, the system becomes hard to test and harder to explain later.
Mode is the execution how
Thinker, Researcher, and future modes belong to task configuration. That lets the same agent be evaluated across execution capabilities without mutating its identity.
Task config is the what
The task config owns horizon, subject scope, input format, output schema, prompt component references, and assembly rules. It is the reusable execution contract.
Components are resolved by reference
Prompt assembly references reusable component identities instead of copying large text everywhere. This supports versioning, controlled upgrades, and clearer Config Details lineage.
System layer
The system layer defines the analytical voice
AI Agent persona
The core identity and investment framework. It answers: who is analyzing this asset and what lens do they use?
Communication tone
The delivery style, such as balanced, expressive, Socratic, investigative, or direct. Tone is not the same as persona identity.
Lexical register
The vocabulary level, such as novice-friendly, intermediate, expert, or slang. This controls readability and domain density.
Content layer
The content layer grounds the forecast in context and data
The active assembly pattern uses a context-enriched structure. Subject context appears before and after global context so the model stays anchored to the specific asset while still considering the macro environment. The final task guidelines then close the prompt by reminding the model to connect macro context back to the subject, forecast horizon, and required output schema.
SYSTEM INSTRUCTION 1. AI Agent persona definition 2. Communication tone 3. Lexical register PROMPT CONTENT 1. Subject context 2. Output instructions and schema requirements 3. Global macro context 4. Subject context repeated for grounding 5. Asset input data and fundamentals where available 6. Task guidelines and required forecast dates
Output instructions
The output format tells the model which structured fields, forecast dates, ratings, scenarios, risks, and timeseries values it must produce.
Global context
Dated macro and world-event context helps the model reason inside the current regime rather than relying only on old training knowledge.
Input data
The asset-specific payload can include latest close price, price statistics, and for equities, financial statement and derived metric context.
Task guidelines
Guidelines remind the model to center the specific asset, use global context carefully, and avoid letting broad narratives override fundamentals.
Sandwiching
Important instructions are repeated around long context
Long-context models often pay strongest attention to the beginning and end of a long prompt, while information buried in the middle can be under-weighted. This is commonly described as the lost-in-the-middle problem. iPulse prompt assembly uses a sandwich-like structure to work with that behavior instead of pretending it does not exist.
Subject before context
The prompt starts by identifying the asset or subject so the model knows what the analysis is about before reading global macro material.
Context in the middle
The global context block can be long, so it sits inside a controlled structure rather than floating as an unanchored wall of text.
Subject repeated after context
Subject grounding is repeated after global context so broad world events do not overpower the asset-specific task.
Guidelines at the end
Task guidelines close the prompt with the scoring, horizon, output, and grounding expectations the model must satisfy.
The same idea is used inside the global context archives themselves: critical summaries and regime takeaways appear near the beginning and are reinforced later, so the most important facts are not stranded in the middle of a long input.
Lineage
Each assembled forecast keeps its configuration identity
Prompt assembly references component IDs and task config IDs so outputs can be associated with the correct agent, mode, tone, register, input format, output format, forecast horizon, and model version. That lineage is what lets iPulse evaluate forecast quality and retire weaker configurations.
In the product, eligible signed-in users can inspect a readable version of this lineage in Prediction Config Details. The modal shows the configuration structure, not a raw full prompt dump.
Why it matters
Good prompts are assembled, not guessed
A simple generic prompt often asks a model to research an asset in isolation. iPulse instead supplies a richer environment: a distinct AI Agent, execution mode, subject-specific data, global context, output schema, and validation expectations. That makes the result more comparable across assets and more useful for educational market research.