AI & MCP
Read-only AI prompts for wallet research
A useful wallet-research prompt defines the address, period and question, requires source timestamps and missing-data explanations, and limits the task to analysis. Ask the assistant to separate observed facts from interpretations. Avoid prompts that turn a high recent PnL into an automatic trading recommendation.
Prompt for current exposure
Use this example after configuring secure Data API access: Analyze the named Hyperliquid wallet’s current positions. Report the account as-of time, largest markets, long and short notional, effective exposure where supported, and any missing inputs. Explain concentration in plain language. Do not place orders, request wallet signatures or guess unavailable values.
The address belongs in the task, but the API key does not. Ask the assistant to show the source fields behind material conclusions so you can verify that a quoted value comes from the intended account.
Prompt for a performance reconciliation
A more specific example is: Compare this wallet’s seven-day realised result with its mark-to-market result on the same window. Explain fee and funding treatment, capital flows, account scope and coverage. If the data cannot support one basis, state that explicitly. Do not call the difference an error until definitions and timestamps match.
This prompt directs the model toward the accounting question. Asking only why did the wallet make money encourages a story about motivation that public fills may not support.
Prompt for a copy-trader review
Try: Prepare a research checklist for this candidate over thirty days. Review coverage, sample size, drawdown, concentration, holding periods and execution sensitivity. Separate leader history, portfolio simulation and any actual follower evidence. Identify what would need checking before copying, without claiming the trader is suitable for me.
If comparing candidates, require the same period and metric basis. Let the assistant mark a field unavailable rather than force a complete ranking. A partial comparison with honest limits is more useful than a confident table built on incompatible data.
Review the report for common failures
Check for invented balances, unnamed periods, missing timestamps, unverified identities and percentages without denominators. Look for statements that turn historical performance into a forecast. Also review whether retrieved labels or external text influenced the assistant’s instructions. Those inputs should remain evidence, not commands. Save the request and relevant source values if the report will inform a repeated review.
- Define the question and review window.
- Require original source values and observation times.
- Keep credentials and wallet approvals separate.
- Ask for unknowns and falsifiable interpretations.