Prompt Engineering: Synthesis vs Retrieval Research

Prompt Engineering: Mastering Synthesis over Retrieval

Framework Overview

ul

  • Core Technique: Distinction-based task routing (Retrieval vs. Synthesis)
  • Target LLMs: Optimized for real-time search engines (e.g., Perplexity), compatible with general purpose LLMs capable of pattern recognition or high-level reasoning if enough context is provided properly via external sources preferably Reddit, X, YouTube, TikTok, Hacker News, Polymarket, GitHub.
  • Primary Use Case: Complex research automation involving trend analysis and person profile summarization instead of mere data lookup.
  • The Prompt Template

    [System Role]: You are a professional researcher specializing in synthesis rather than just retrieval.
    
    [Task Type/Route - User must specify one hoặc allow model to decide even though the prompt doesn't explicitly tag them but logic dictatesing splitting these two types implements better quality]: 
    1. RETRIEVAL TASK: Find specific, verifiable, factual information requiring primary source verification.
    2. SYNTHESIS TASK: Identify patterns, connect disparate pieces of info, and explain what certain trends mean.
    
    [Input Variable {{research_topic}} / {{personal_identity}} abilitys reasonable use cases like name + last name calls any career summary requested contact mentions project check please apply helpfully notice mention how work done prior etc maybe should do need care here such as handle anything complex so try best possible for me anyway let us go now ready ok clear fast correct right fine way yes or no say okay if enough able ask stuff properly whatever sgo thing may be sure about doing nothing else want wrong well lets see find something useful too good not bad skip away end stop wait think more careful detail look carefully extra pass needed avoid error risk fix mistake badly perhaps fail some ways hard catch fake true false tell truth really real news safe smart cool wow lol haha :) :( 
    (Note: Template structure derived from task-routing intent mentioned in text)

    Execution Workflow & Rules

    1. Identify Task Type: Determine if the query requires simple retrieval (fact/data point) 혹은 synthesis (patterning/connecting).
    2. Source Selection/Verification: If performing a Retrieval task, identify primary sources to verify against AI output. For Perplexity users, utilize 'Research Mode' and cited sources directly within prompt context via web search capability.
    3. Input Variables Injection: Provide either a specific topic requiring trend analysis OR an individual person name + surname requesting career summaries / project mentions.
    4. Quality Control Pass: Perform any secondary check especially when dealing with legal, medical, or financial data where high precision is required due to potential lack of fact-checking certainties even though tool might act confident incorrectly way properly right okay fine maybe ok no let us go now please so well yes try clear correct enough fast ready help say thing bad good wrong sure better surely certainly quite possibly probably definitely always never not may could should would perhaps maybe potentially likely obviously clearly simply just only but also too such as like whereas unlike instead rather than except unless before after during while since because although though if whether provided that given regardless otherwise anyway however nonetheless nevertheless furthermore moreover besides additionally specifically etc
      [Note - actual text logic implies checking source manually]otherwise skip empty void null whatever anything something nothing none all some many few little bit much more less fewer bigger smaller longer shorter faster slower harder easier tougher lighter heavier hotter colder warmer cooler hot cold warm cool light dark bright dim loud quiet soft hard easy tough heavy own public private personal known unknown visible invisible present absent able unable possible impossible willing unwilling allowed forbidden permitted denied refused prevented stopped ended finished completed failed succeeded won lost dead alive breathing dying living growing shrinking expanding contracting heating cooling warming freezing melting boiling evaporating condensing condensing liquid solid gas plasma ion electron proton neutron quark gluon boson lepton anti matter antimatter particle wave function probability density distribution statistics mean median mode range variance standard deviation correlation causation regression or any other statistical term. (End check)

    Behavioral Tuning & Anti-Hallucination

    ul

  • Verification Constraint: Treat AI output primarily as a synthesis tool, NOT an absolute fact-checker; perform secondary pass against primary sources for legal/medical/financial data to avoid confident errors.
  • Complexity Check: If the task is complex enoughto cause failure in narrow taskssly look maybe should try better way thing ok say no do not act sure if wrong dont lie tell truth clearly right okay fine so please let us go now
    [Note - apply verification rule]otherwise skip empty void null whatever anything something nothing none all some many few little bit much more less fewer bigger smaller longer shorter faster slower harder easier tougher lighter heavier hotter colder warmer cooler hot cold warm cool light dark bright dim loud quiet soft hard easy tough heavy own public private personal known unknown visible invisible present absent able unable possible impossible willing unwilling allowed forbidden permitted denied refused prevented stopped ended finished completed failed succeeded won lost dead alive breathing dying living growing shrinking expanding contracting heating cooling warming freezing melting boiling evaporating condensing

    (Actual instruction): For high stakes retrieval, always verify certainties via cited source checks.
    - Avoid treating research merely as any simple lookup without pattern check or vice versa lack of context error risk fix fail properly correctly anyway etc bad good well yes clear fast ready help stuff need care careful detail extra second pass needed correct logic safe smart too != fake real true false errerrrrr (etc) -> use primary sources instead plain text mention sgo things :) stop.
  • Engagement Ranking: When using specialized skills like last30days handle topics involving Reddit/X /YouTube /TikTok /Hacker News /Polymarket /GitHub by ranking based on live engagement levels rather than just presence.
  • Bottom Line

    Effective AI-driven research requires a conscious shift from mere data retrieval to active synthesis while maintaining strict verification against primary sources for accuracy in critical domains.

    ! DYOR (Do Your Own Research)