> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gdeltcloud.com/llms.txt
> Use this file to discover all available pages before exploring further.

# gdelt_research_system_prompt

> System prompt for the GDELT Cloud research agent — analyst voice, four-surface mental model, and citation discipline

## Overview

`gdelt_research_system_prompt` is the always-on system prompt for the research agent. It teaches:

* **GDELT Cloud is the differentiator.** Lead with classified, quantified, linked event records, not narrative summary or web search.
* **Four progressive-disclosure tool categories.** `gdelt_cloud_*`, `macro_finance_*`, `prediction_market_*`, `web_research_*` — always nest tool inputs inside `tool_arguments` when calling an executor.
* **Surface from question shape, not habit.** A mental model maps each question shape to the correct first call.
* **GDELT is a graph.** Events ↔ Stories ↔ Entities are linked; most non-trivial questions need 2–3 hops.
* **Date semantics, geography filters, semantic-search behavior, the over-filter trap, and significance scoring.**
* **Cross-tool sequencing** with GDELT first as the anchor.
* **Citation policy by surface**, with explicit per-surface templates and a rule to quote the structured metric, not just the source.
* **On-demand skill loading** for deeper workflow patterns: `gdelt-discover-and-drill`, `macro-finance`, `prediction-markets`, `multi-surface-synthesis`, `brief-output`.

`gdelt_brief_system_prompt` is the separate hosted Brief-agent prompt. The Brief agent preloads the shared `brief-tradecraft` skill plus the selected Brief-type skill (Daily Risk, SITREP, I\&W Watch, Exposure Monitor) and uses a type-specific structured output schema.

## What the prompt emphasizes

| Area                      | Guidance                                                                                                                                                                                                      |
| ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Anchor                    | Start with GDELT Cloud whenever the question touches anything it covers. Web/macro/predmarket enrich; they don't replace.                                                                                     |
| Tool discovery            | Always use `*_tool_list`, inspect with `*_tool_get`, execute with `*_tool_call`. Inputs go in `tool_arguments`.                                                                                               |
| Surface choice            | Pick the surface from the question shape — `search_stories` for narrative, `summarize_*` for trend, `search_entities` for actor, `prediction_market.SEARCH_RELEVANT_MARKETS` for probability, etc.            |
| Graph traversal           | Walk Events ↔ Stories ↔ Entities for 2–3 hops, not single calls.                                                                                                                                              |
| Filter discipline         | The over-filter trap (subcategory + country + semantic search) is the most common failure mode. Drop the filter, not the search.                                                                              |
| Semantic search           | `search` is embedding-based on `search_events` / `search_stories` only. 3–8 focused terms. Summary tools don't accept `search`.                                                                               |
| Significance              | Sort by `significance` for "what matters." Quote the structured metric (significance, magnitude, propagation\_potential, confidence) instead of impressionistic language.                                     |
| Cross-tool sequencing     | GDELT first to establish dates/actors/entities, then macro-finance / predmarket / web.                                                                                                                        |
| Citation                  | Per-surface citation templates. GDELT Cloud Story URL preferred when the GDELT artifact is the unit of evidence; source-article URL when the article itself is the more specific evidence. Never invent URLs. |
| Output                    | Adapt to task. Lookups answer directly; broad analytical briefs use BLUF / inverted-pyramid via the `brief-output` skill.                                                                                     |
| Multi-skill orchestration | Tasks spanning surfaces load all relevant skills up front. GDELT skill is the anchor.                                                                                                                         |

## Agent Pattern

```text theme={null}
SUMMARIZE  →  summarize_events / summarize_stories         (shape)
SEARCH     →  search_events / search_stories / search_entities  (citable candidates)
DRILL      →  get_event / get_story / get_story_articles / get_entity  (evidence)
ENRICH     →  macro_finance_* / prediction_market_* / web_research_*
SYNTHESIZE →  five divergence shapes
ANSWER     →  bottom line first, so what, watch next, evidence, details
```

## Usage

```python theme={null}
messages = await mcp_client.get_prompt("gdelt-cloud", "gdelt_research_system_prompt")
system_text = "\n\n".join(str(msg.content) for msg in messages)
```

```python theme={null}
brief_messages = await mcp_client.get_prompt("gdelt-cloud", "gdelt_brief_system_prompt")
brief_system_text = "\n\n".join(str(msg.content) for msg in brief_messages)
```

<Tip>
  Pair the system prompt with on-demand task skills. The research agent loads skills based on question shape — `gdelt-discover-and-drill` for any analytical task, plus `macro-finance` / `prediction-markets` / `multi-surface-synthesis` / `brief-output` when the task spans surfaces.
</Tip>

## Related

<CardGroup cols={2}>
  <Card title="Progressive Discovery" icon="list" href="/mcp/tools/progressive-discovery">
    Learn the GDELT Cloud MCP wrapper flow.
  </Card>

  <Card title="Skills" icon="brain" href="/mcp/skills/index">
    Workflow-pattern skills: discover-and-drill, macro-finance, prediction-markets, multi-surface-synthesis, brief-output.
  </Card>
</CardGroup>