Skip to main content
Once the server is connected, you talk to your assistant normally, and it decides which Blockworks tools to call. The prompts below are patterns that work well, grouped by what you are trying to do. Swap in the assets, protocols, and timeframes you care about. If you want to know which tool a given prompt is likely to reach for (document search, the tabular pair, or the timeseries pair), see the tools reference.

​
Market data lookups

Point lookups where you want a current number, not a narrative.

​
Protocol and network research

Questions that need context assembled from several data sources.

​
Screening and comparison

Ask for a set rather than a single fact, and say what to rank on.

​
Monitoring and news

Catch-up prompts and recurring checks.

​
Getting better answers

For complex or multi-part questions, ask several focused questions instead of one loaded one. Single-topic requests usually come back with more depth and fewer gaps.
Quantitative questions work best when you state:
  • The metric: “TVL”, “market cap”, “24h trading volume”, “daily active addresses”.
  • The entity: the asset, protocol, or network by name.
  • The timeframe: “last 30 days”, “YTD”, “since the merge”.
  • The baseline, if you are comparing: “vs last month”, “vs Ethereum”.
A few habits that pay off:
  • Ask for the shape of the output. “As a table”, “as a bulleted summary”, “with the raw numbers” all work.
  • Ask what is available first. “What Blockworks datasets cover DeFi protocols?” makes the assistant read the catalog, which tells you what is queryable before you write a more specific prompt.
  • Name concrete dates for time-sensitive questions. “The week of March 10” retrieves a sharper set of documents than “latest”.
  • Give research questions time. Dataset reads come back quickly, but a document search runs an agent over the corpus and can take a while; a slow answer is usually still on its way.
  • Ask it to cite what it pulled. Requesting the source or the underlying figures makes it easy to spot when an answer came from the model’s own knowledge instead of from Blockworks data.
  • Follow up rather than restart. The assistant keeps the earlier results in context, so “now break that out by chain” is cheaper and more accurate than a fresh, longer prompt.