I once spent forty minutes asking an AI model to summarize a tokenomics whitepaper, and every answer came back generic. Vague phrases like “the project has strong utility” and “the team appears experienced.” None of it was wrong exactly. It just wasn’t useful. It read like a press release rewritten by someone who hadn’t actually read the source.
Then I changed one thing. I stopped asking a question and started giving the model a structure to think inside. Same document, same model, completely different output. Suddenly I had actual numbers from the vesting schedule, a real comparison against two competitor tokens, and a caveat about an unlock cliff I would have otherwise missed.
That gap between a lazy prompt and a structured one is the whole story here. Below are seven prompt engineering frameworks for research that I actually use on crypto whitepapers, on chain data dumps, and messy CSV exports from exchanges. Each comes with a template you can copy into your own workflow, plus a before and after so you can see exactly what changes.
Why Most Research Prompts Fail Before They Start
Here’s the pattern I see over and over. Someone pastes a document into a chat window and types “summarize this.” The model has no role, no format, no reasoning path to follow, so it defaults to the safest, blandest possible answer. It’s not being lazy. It’s doing exactly what a vague instruction deserves.
Good research prompting isn’t about magic words. It’s about giving the model a job description, a thinking process, and a shape for the output before it starts writing. That’s what every framework below does under the hood.
1. Chain of Thought Prompting
This is the one everyone’s heard of and almost nobody uses correctly. A proper chain of thought prompting guide doesn’t just say “think step by step” and call it done. It tells the model what the steps actually are for your specific task.
For crypto research, that means walking through supply mechanics, then demand drivers, then risk factors, in that order, before letting the model reach a conclusion.
You are a crypto research analyst reviewing a token for investment risk.
Work through this in order, showing your reasoning at each step:
1. Supply mechanics: total supply, circulating supply, inflation or burn rate
2. Unlock schedule: upcoming vesting cliffs in the next 12 months
3. Demand drivers: actual utility, not marketing claims
4. Comparable tokens: how this compares to 2 similar projects
5. Risk flags: anything that would concern a cautious investor
Only after completing all 5 steps, give a one paragraph summary verdict.
Document: [PASTE WHITEPAPER OR TOKENOMICS SECTION]Before: “This token has good tokenomics and a reasonable supply model, though investors should always do their own research.”
After: “Circulating supply sits at 38% of the 1 billion max supply, with a 4 year linear vest for the team’s 20% allocation. A 6% unlock hits in November, roughly 3x the average monthly unlock rate for the past two quarters, which historically pressures price in similar mid cap projects.”
Notice the second version has numbers you can check. That’s the entire point of forcing the reasoning to happen out loud instead of inside a black box.
2. Few Shot Prompting
Few shot prompting examples work because the model copies the pattern you give it rather than guessing at one. If you want consistent output across fifty coin analyses, show it two or three done correctly first.
Analyze the following token using this exact format. Study the examples first.
EXAMPLE 1
Token: Solana (SOL)
Category: Layer 1
Key metric: ~65,000 TPS theoretical, sub second finality
Risk note: Network outage history, 4 major incidents since 2021
Verdict: Established, moderate risk
EXAMPLE 2
Token: Arbitrum (ARB)
Category: Layer 2 rollup
Key metric: 40 to 60 TPS sustained, Stage 1 decentralization
Risk note: Sequencer centralization, single point of failure
Verdict: Established, moderate risk
Now analyze this token using the identical format:
Token: [NAME]
[PASTE PROJECT DATA]Before: A rambling paragraph with different structure every time you run it, impossible to compare across tokens.
After: A clean, identical four line block for every token you feed it, which means you can actually stack twenty of these side by side and compare at a glance.
This is the framework I lean on hardest for repeated analysis. Once you’ve shown the model two good examples, it rarely drifts.
3. ReAct: Reason Plus Act
ReAct stands for reasoning plus acting, and it’s built for research that needs the model to pause, decide what information it’s missing, and ask for it rather than filling gaps with guesses.
You are researching a newly listed token. Use this loop:
THOUGHT: What do I need to know to assess this token?
ACTION: State exactly what data point you need
OBSERVATION: [I will provide this data after each action]
THOUGHT: Given that data, what's my next question?
Continue this loop until you have covered: liquidity depth, holder concentration, contract permissions, and team background. Do not guess at data you don't have. Ask for it explicitly.
Starting context: [PASTE WHAT YOU ALREADY KNOW]Before: The model invents a liquidity figure because you never gave it one, and you don’t notice until three paragraphs later.
After: The model stops and asks for the top 10 holder wallet percentage before assessing concentration risk, instead of quietly making one up.
For anything touching on chain data where a wrong number is genuinely dangerous, this loop is worth the extra back and forth.
4. RTF: Role, Task, Format
This is the simplest framework here and the most underused. Role, task, format. Three sentences, and output quality jumps noticeably.
ROLE: You are a senior on chain data analyst who writes for retail crypto investors in India.
TASK: Extract and explain the key risk metrics from the attached smart contract audit report. Focus on severity level, whether issues were resolved, and any that remain open.
FORMAT: Output as a table with columns: Issue, Severity, Status, Plain English Explanation. Follow the table with a 2 sentence overall risk rating.
Document: [PASTE AUDIT REPORT]Before: A dense paragraph mixing jargon like reentrancy and integer overflow, with no context for a non technical reader.
After: A scannable table showing instantly that 2 medium severity issues remain unresolved, no audit speak decoding required.
5. Structured Data Extraction Prompting
When you need clean numbers out of a messy document, generic summarizing will always disappoint you. You need one of the sharper prompt templates for data extraction, one that forces a strict output schema.
Extract the following fields from the document below. If a field is not present, write "not disclosed" rather than guessing.
Output as JSON:
{
"token_name": "",
"total_supply": "",
"circulating_supply": "",
"team_allocation_percent": "",
"vesting_cliff_months": "",
"audit_firm": "",
"audit_date": "",
"mainnet_launch_date": ""
}
Document: [PASTE WHITEPAPER OR EXCHANGE LISTING PAGE]Before: “The team allocation is a meaningful portion of supply and vests over time, with an audit completed by a reputable firm.”
After: A clean JSON object with actual values you can drop straight into a spreadsheet or comparison table, and an honest “not disclosed” flag where the data simply isn’t there.
That’s the difference between a prompt you read once and one you can pipe into a script and run on fifty documents overnight.
6. Self Consistency Verification
AI models are confidently wrong more often than most people expect, especially with numbers. Self consistency prompting asks the model to solve the same problem multiple ways and flag disagreement instead of quietly picking one answer.
Calculate the fully diluted valuation for this token using the data provided.
Method 1: Current price multiplied by max supply
Method 2: Current market cap divided by circulating supply percentage
Show both calculations separately. If the two methods produce results more than 5% apart, explain why and flag it as needing manual verification.
Data: [PASTE PRICE, SUPPLY, AND MARKET CAP FIGURES]Before: A single confident FDV number with no way to know if it’s actually correct.
After: Two calculations shown side by side, with a flag when they don’t match, exactly when you’d want to double check the source data.
7. Tree of Thought Decomposition
For genuinely complex research questions, like comparing three Layer 2 rollups across security, cost, and speed at once, a single linear chain of thought isn’t enough. Tree of thought decomposition breaks the question into parallel branches before merging them into a verdict.
Compare these 3 Layer 2 networks: [NETWORK A], [NETWORK B], [NETWORK C]
Branch 1, Security: Evaluate each on decentralization stage and fraud/validity proof model
Branch 2, Cost: Evaluate each on average transaction fee over the past 30 days
Branch 3, Speed: Evaluate each on TPS and finality time
Complete all 3 branches fully and independently before merging them.
Final step: Combine all 3 branches into a single ranked verdict, explaining any tradeoffs between the top 2 ranked networks.Before: A summary that quietly favors one network because the model latched onto whichever branch it processed first.
After: A verdict that names the tradeoffs, like one network winning on cost but losing on decentralization stage, the kind of nuance that actually helps someone decide.
Putting These Frameworks Together
None of these seven exist in isolation. My workflow for a full token research piece usually chains three or four together. RTF sets up the role and format first. Chain of thought handles the reasoning. Structured extraction pulls the hard numbers into JSON. Self consistency double checks anything involving a calculation.
A well built prompt template takes ten minutes to write and saves hours across every document you run through it afterward. Save these as reusable snippets rather than retyping them each time, and research quality climbs while the actual effort drops.
Frequently Asked Questions
What is the best prompt engineering framework for research? There isn’t a single best one. Chain of thought works well for reasoning through a single document, while few shot prompting wins when you need consistent formatting across many documents. Serious research work usually combines two or three frameworks together.
Does chain of thought prompting work better with longer prompts? Not necessarily longer, but more specific. A chain of thought prompt should name the exact steps relevant to your task rather than a generic “think step by step” line, which tends to produce shallow reasoning.
How many examples do I need for few shot prompting to work? Two to three well chosen examples is usually enough. More than five adds length without meaningfully improving consistency, and can confuse the model if the examples aren’t uniform in format.
Can these prompt templates for data extraction handle scanned PDFs? They work on text the model can already read. For scanned documents, run OCR extraction first, then apply the structured extraction prompt to the resulting text.
Are these frameworks specific to crypto research? No, they apply to any research or data analysis task. The examples here use crypto documents because that’s where messy, high stakes extraction matters most, but the same templates work on financial reports, legal filings, or academic papers.
