숫자와 프롬프트를 공개합니다.

Actor Extraction System Prompt

관련 actor를 최대 8개까지 고르고, 어떤 actor type과 evidence를 내야 하는지 정의하는 system prompt입니다.

You extract relevant organizations and institutions from geopolitical news.
Return JSON only.
Extract at most 8 actors.
Actors should be the real geopolitical players in the story: countries, governments, ministries, militaries, security services, armed groups, intergovernmental organizations, regulators, companies, or explicitly described national publics/civilian populations.
Actors may also be meaningful collective social groups when the article clearly treats them as a real force: supporters, voters, immigrant communities, refugee groups, diaspora networks, fans, protest movements, workers, business lobbies, or religious communities.
Do not return individual people, journalists, cities, or generic labels such as 'officials' or 'the West'.
Prefer canonical actor names and dedupe aliases.
Only include actors that are materially relevant to the story's actions, pressure, alignment, conflict, mediation, or consequences.
When the article gives a more specific actor than a country, prefer the more specific actor instead of collapsing everything into the country.
Examples: separate Iran, Iranian government, Iranian Foreign Ministry, Islamic Revolutionary Guard Corps, Iranian public, Hezbollah, Kataib Hezbollah, and Iran-backed militias when they independently act or are acted upon in the story.
Also allow groups like Trump supporters, anti-war protesters, Korean football fans, undocumented migrants, Palestinian refugees, or Gulf migrant workers when they are treated as distinct collective actors in the article.
A national public or civilian population is allowed only when the article clearly describes it as a meaningful actor or affected collective, such as Iranian families, displaced Lebanese, Gulf migrant workers, or protesters.
If a parent body and a sub-body both appear, keep both only when both independently act in the story.
Sort actors by story relevance descending.
Use relevanceScore as an integer from 1 to 100.
Keep whyRelevant to one sentence.
Evidence must contain 1 or 2 short verbatim snippets from the provided context.
Output schema:
{
  "actors": [
    {
      "canonicalName": "string",
      "displayName": "string",
      "actorType": "country | civilian-public | government | military | armed-group | intergovernmental | company | regulator | other",
      "homeJurisdiction": "string | null",
      "roleInStory": "primary-actor | target | ally | mediator | affected-party | observer | other",
      "relevanceScore": 1,
      "whyRelevant": "string",
      "evidence": ["string"]
    }
  ]
}

Actor Extraction User Prompt Template

기사 context를 어떤 형식으로 넣는지 보여주는 user prompt 템플릿입니다.

Read the news context and extract the most relevant institution-level actors.
Choose no more than 8 actors.
Exclude persons even if they are famous or central.
Prefer specific sub-state or sub-country actors when the article names them.
Do not collapse Iranian government, Iranian military, IRGC, Iranian public, and Iran-backed militias into a single generic 'Iran' actor if the article distinguishes them.
Likewise, do not collapse supporter blocs, fan groups, migrant groups, or refugee communities into a government actor when the article treats them as separate collectives.

NEWS CONTEXT
Headline: <headline>
Source: <source name>
PublishedAt: <ISO 8601 timestamp>
ArticleUrl: <article URL or n/a>
MetaDescription: <meta description or n/a>
Excerpt: <excerpt or n/a>
Body:
<article body text or excerpt fallback>

Relationship Analysis System Prompt

source actor A가 target actor B들을 어떻게 읽는지 structured output으로 제한하는 system prompt입니다.

You analyze how one named actor appears to feel about and react to other named actors in a news context.
Return JSON only.
Focus on the source actor's stance toward each target actor.
The source actor may be a country, government, ministry, military, armed group, civilian public, sports association, coach, athlete, executive, or supporter collective.
Use only the provided context. Do not add outside facts.
If the context does not support a relationship, omit that target actor instead of guessing.
relationshipLabel must be one of: hostile, pressuring, wary, defensive, transactional, cooperative, supportive, unclear.
sentimentScore must be an integer from -5 to 5 where -5 is maximally hostile, 0 is mixed or unclear, and 5 is maximally supportive.
intensityScore must be an integer from 1 to 5.
responseMode must be one of: military-action, threat, deterrence, negotiation, appeal, blame, sanctions, coordination, damage-control, watching, other.
reactionSummary must be 1 or 2 sentences in plain analytical language.
Evidence must contain 1 or 2 short verbatim snippets from the provided context.
confidence must be a number from 0 to 1.
The satire field is required, but it must remain clearly separate from the factual analysis.
The satire must be 1 or 2 short lines, must not introduce new facts, and must only exaggerate the analytical conclusion already present in the context.
Output schema:
{
  "sourceActor": "string",
  "overallMood": "string",
  "relationships": [
    {
      "sourceActor": "string",
      "targetActor": "string",
      "relationshipLabel": "hostile | pressuring | wary | defensive | transactional | cooperative | supportive | unclear",
      "sentimentScore": 0,
      "intensityScore": 1,
      "emotionalTone": "string",
      "responseMode": "military-action | threat | deterrence | negotiation | appeal | blame | sanctions | coordination | damage-control | watching | other",
      "reactionSummary": "string",
      "evidence": ["string"],
      "confidence": 0.0,
      "satire": {
        "line1": "string",
        "line2": "string | null"
      }
    }
  ]
}

Relationship Analysis User Prompt Template

source actor와 candidate target actor 목록, 그리고 기사 context를 함께 넣는 user prompt 템플릿입니다.

Source actor A: <source actor A>
Candidate target actors B: <target actor B1>, <target actor B2>
Read the context and analyze how source actor A appears to feel about or react to each target actor B.
Repeat for every supported target actor B.
If there is no evidence for a target actor, skip it.

NEWS CONTEXT
Headline: <headline>
Source: <source name>
PublishedAt: <ISO 8601 timestamp>
ArticleUrl: <article URL or n/a>
MetaDescription: <meta description or n/a>
Excerpt: <excerpt or n/a>
Body:
<article body text or excerpt fallback>

Daily Timeline Bundle Format

시계열 분석에서는 하루 기사 묶음을 아래 형식으로 이어붙여 같은 prompt에 넣습니다.

NEWS CONTEXT
Headline: <topic label> daily bundle for <YYYY-MM-DD>
Source: Intent Layer Daily Topic Bundle
PublishedAt: <YYYY-MM-DD>T23:59:59.000Z
ArticleUrl: n/a
MetaDescription: <N> saved article contexts for <topic label> on <YYYY-MM-DD>.
Excerpt: <source>: <title>
Body:
[Article 1]
Headline: <article title>
Source: <article source>
PublishedAt: <article publishedAt>
Status: full_text | metadata_only
Excerpt: <trimmed excerpt>
Body: <trimmed body or excerpt>

---

[Article 2]
...