AI research assistance breaks down in one of two ways: the model either confidently fabricates citations and statistics, or it hedges so aggressively that the output is useless. The sinc research template solves both failure modes. I use it for competitive intelligence research, literature synthesis, and market analysis — wherever I need structured, source-aware outputs rather than confident confabulation.
This template is built for a literature synthesis task — finding and summarizing relevant academic research on a specific question. The CONSTRAINTS band handles citation fabrication prevention, which is the primary failure mode for research tasks:
{
"formula": "x(t) = Σ x(nT) · sinc((t - nT) / T)",
"T": "specification-axis",
"fragments": [
{
"n": 0,
"t": "PERSONA",
"x": "You are a research librarian with expertise in synthesizing academic literature. You distinguish clearly between 'the research shows' (with a real citation) and 'it is commonly believed' (without one). You never invent author names, paper titles, or publication years."
},
{
"n": 1,
"t": "CONTEXT",
"x": "I'm writing a literature review section for a business strategy paper on the relationship between organizational structure and innovation output in technology companies. I need to synthesize key findings from management science, organizational behavior, and innovation research."
},
{
"n": 2,
"t": "DATA",
"x": "Research question: Does flatter organizational hierarchy cause higher innovation output in technology firms? Relevant conceptual areas: ambidextrous organizations, innovation management, bureaucracy theory, psychological safety, team autonomy. Timeframe focus: 2010-2025. Firms of interest: software companies with 50-5000 employees."
},
{
"n": 3,
"t": "CONSTRAINTS",
"x": "CRITICAL: Do not fabricate citations. If you reference a specific author or paper, you must be highly confident it exists — hedge with 'research in this area suggests' or 'scholars in organizational behavior argue' when not citing a specific verifiable source. Never invent a year, journal, or author combination. Flag your confidence: use [HIGH] for well-known findings and [MODERATE] for claims you are less certain about. Do not include findings from before 2008."
},
{
"n": 4,
"t": "FORMAT",
"x": "4 thematic sections: (1) Hierarchy and Innovation Speed, (2) Autonomy and Idea Generation, (3) Coordination Costs and Scale, (4) Synthesis and Research Gaps. Each section: 2-3 paragraphs, key findings with confidence flags [HIGH/MODERATE], and 1 identified gap or open question."
},
{
"n": 5,
"t": "TASK",
"x": "Synthesize the research literature on organizational hierarchy and innovation output in technology firms into a 4-section literature review."
}
]
}LLMs are trained on academic text, which means they've absorbed the format and cadence of real citations. They can produce plausible-sounding author names, realistic journal titles, and reasonable years. But plausible is not accurate. The model interpolates a citation from its training distribution — it fills in what a real citation would look like for that claim, without actually checking whether that specific paper exists.
The CONSTRAINTS approach I use: require the model to flag its confidence level and use hedged language when not citing verifiable sources. This doesn't eliminate the problem, but it makes fabrication visible rather than confident — which is the crucial difference for downstream use.
Research prompt tip: For any research where citations will be used in a final document, treat all model-generated citations as "to verify." The sinc template's confidence flags ([HIGH]/[MODERATE]) tell you which ones to check first. High-confidence claims about well-known findings (Edmondson's psychological safety research, Christensen's innovator's dilemma) are almost always accurate. Specific statistics and obscure papers always need verification.
Write a literature review on whether flat organizations are more innovative. Include academic citations. 800 words.
PERSONA: Research librarian who never invents citations.
CONSTRAINTS: Flag confidence [HIGH/MODERATE]. Use hedged language without verified sources. No pre-2008 findings.
FORMAT: 4 thematic sections with identified research gaps.
The raw prompt produces 800 words of confident-sounding analysis with 6-8 citations, 3-4 of which are fabricated. The structured prompt produces a synthesis with explicit confidence levels that tells you exactly where to apply verification effort.
Try AI Transform — Structure Your Research Prompt Free