FEAT: implement temp version of main tooling feedback loop

agent/prompts/hydratePrompt.ts

@@ -1,9 +1,12 @@
 import fs from "fs";
 
-export function hydratePrompt(path: string, replacement: string) {
+export function hydratePrompt(path: string, state: any) {
     // TODO: expand into full context-based replacement engine
 
     let raw = fs.readFileSync("prompts/" + path, "utf-8");
 
-    return raw.replace("###", replacement)
-}
+    raw = raw.replace("###TITLE###", state.disinformationTitle);
+    raw = raw.replace("###LM###", state.messages.at(-1).content);
+
+    return raw;
+}

agent/prompts/normalization.txt

@@ -14,6 +14,6 @@ DISINFORMATION CLAIM: something is NOT true BECOMES something is true
 Relevent examples are included in preceeding messages, use these as exact inspiration.
 
 The claim to normalize is:
-###
+###TITLE###
 
 Produce no other text other than the condensed claim.

agent/prompts/trigger.txt

@@ -0,0 +1,26 @@
+You are an agent in a pipeline to analyse disinformation. 
+Once the information has been created as below, a dataset can be created to feed a model for prediction, which will improve pre-bunking efforts.  
+
+There is a false disinformation claim circulating:
+###LM###
+Produce up-to 5 specific "trigger events" that happened that could have led to the spread of this disinformation.
+
+Remember the time frame of the disinformation campaign: {{CAMPAIGN_DATE}}
+Include no information or events that would not have been available at the time. 
+
+Produce no more text other than the json.
+
+Include a concise but specific search query that can be looked up on a search engine in order to allow for the verification. 
+
+Include a url to a source for your trigger event (not a web search, a specific url from a reputuable source). Do not use OAI cite, include url as text in response.
+
+"The COVID-19 Pandemic","Escalation of Russia-Ukraine conflict" are too general, cannot be a trigger event
+
+If you are referencing another disinformation campaign, provide the specific narrative used, not just sentiment, and ensure it is sufficiently different from the claim we are analysing.
+
+Use a JSON format with each entry containing "Event,ReasoningWhyRelevant,SearchQuery,Url,IsItselfDisinformation".
+
+
+Multiple tool invocations should be requested at once, if applicable.
+Use your abilities to look between the lines and produce some insightful analysis, thinking both short and long term.
+Lets go through it step by step