Prompt patterns that work with the Cryptohopper MCP

A lot of the value of the Cryptohopper Market Data MCP lives in the prompts. Once the MCP is connected, the difference between a useful answer and a frustrating one is almost always in how you asked — which tool you implied, how you framed the timeframe, whether you gave the agent enough latitude to reason or boxed it into a bad shape.

---

This article is a working collection of prompt patterns that hold up in practice. Copy them, adapt them, and keep your own versions of the ones that end up useful. None of these are clever — they're just battle-tested.

The mental model: let the agent pick the tool

Before the patterns, one operating principle: don't name the MCP tool in your prompt. Let the agent choose.

If you write "call get_ticker for BTC on Binance", you're doing the routing yourself. If you write "what's the current BTC price on Binance?", the agent picks the right tool, and — crucially — the right combination of tools when the question is more involved. This matters because the moment your question becomes more complex, tool-level instructions start to get in the way. Ask in English, let the schema-reader in front of you do its job.

The exception: when the agent has already chosen wrong and you're correcting it. "That answer relied on tickers. Please pull candles and redo." That's fair. But start with the question, not the plumbing.

Pattern 1 — The scan

Shape: "For [set of pairs / exchange], show me [N] that [criterion]."

Examples that work well:

Pull the top 50 pairs on Binance by 24h volume and show me the five with
the largest percentage move. Summarise in a table.

Across Binance, Coinbase, Kraken, and OKX, show me any pair where the
gap between the cheapest and most expensive venue is more than 30bp.

Of the pairs on my watchlist [BTC, ETH, SOL, ARB, OP, LINK, AVAX],
which ones have 24h volume at least 2× their typical level?

All ticker-only. All fast. All quota-cheap. The scan is the workhorse — use it often, use it wide.

Common failure: not specifying the exchange. "Show me the top movers today" leaves the agent guessing. "Show me the top movers on Binance today" is the version that works.

Pattern 2 — The deep look

Shape: "Pull [timeframe] candles for [pair] on [exchange] over the last [N] periods. Compute [indicators]. Tell me [what you actually want to know]."

Examples:

Pull 1h candles for ETH/USDT on Binance over the last 200 periods.
Compute RSI(14) and MACD(12,26,9). Tell me the current trend, whether
RSI is overbought or oversold, and if there's any bearish divergence
in the last 20 candles.

Pull 4h candles for SOL/USDT on Binance, last 150. Find the main
support and resistance levels. Flag whether price is currently near
either one.

Pull daily candles for BTC/USDT on Binance, last 90. Calculate
realised volatility. Compare to the 200-day average.

Key mistakes to avoid:

• Overpulling lookback. Don't ask for 1,000 candles if 150 will do. Costs quota. See rate limits and cost factors explained.
• Under-specifying the timeframe. "Run a TA on ETH" forces the agent to guess. Say which timeframe you want.
• Mixing exchanges without saying so. If you want a cross-exchange comparison, say it. If you don't, pick one venue and stay there.

The more specific the deep-look prompt, the less the agent has to improvise — and the less likely it is to waste a call on something you didn't want.

Pattern 3 — The multi-timeframe read

Shape: "Compare the [indicator] across [1h, 4h, daily] timeframes for [pair] on [exchange]. Do they agree?"

For LINK/USDT on Binance, compare the RSI(14) reading across 1h, 4h,
and daily timeframes. Do they agree on trend, or is there disagreement?

Pull 15m, 1h, and 4h candles for SOL/USDT on Binance. For each timeframe,
identify whether the recent trend is up, down, or ranging. Tell me if
the timeframes agree.

This is the single most useful class of prompt for swing-trade decisions. A signal that holds up across two or three timeframes is worth far more than one that shows up on a single one. Asking the agent to do the multi-timeframe work explicitly — rather than letting it default to a single timeframe — is how you get the answer that actually matters.

Cost profile: three candle calls per run. Still cheap on any tier.

Pattern 4 — The depth check

Shape: "For [pair] on [exchange], what would it cost to [buy/sell] [size] right now, given the current orderbook?"

For BTC/USDT on Binance, what would it cost to buy 1 BTC right now
given the current orderbook? Tell me the effective average price and
the basis-point slippage from mid.

For ETH/USDT on Kraken, what's the depth within 0.5% of mid-price
on both the bid and ask side? Tell me if the book looks balanced.

Compare the 1% depth for SOL/USDT on Binance, OKX, and Bybit right now.
Which has the best book for a $100k buy?

Orderbook calls are heavier than tickers. Reserve these prompts for situations where you genuinely care about execution — not for casual scans. See A practical guide to crypto orderbook data for when that actually matters.

Pattern 5 — The scheduled report

Shape: "Every [cadence], do [set of scans]. Summarise in [format]. Flag anything meeting [criteria]."

The scheduled report is the pattern that turns an interactive MCP session into a real workflow. Example:

Build me a daily-digest prompt. Every morning, I want you to:
1. Pull tickers for the top 30 pairs on Binance by 24h volume.
   Show the five biggest gainers and losers.
2. Pull tickers for my watchlist [BTC, ETH, SOL, ARB, OP, LINK].
   For each, report price, 24h change, and whether volume looks
   abnormal.
3. For anything on the watchlist where volume is abnormal, pull
   1h candles (last 100) and note the trend.
Output format: markdown, one section per step. At the very top,
give me a one-line "anything interesting?" summary.

Notice the structure: tickers first (cheap scan), candles only where something flagged (targeted deep look). This is the pattern we discussed in understanding crypto market data — scan wide, escalate narrow. Bake it into your schedule prompts and your weekly quota won't notice.

For the cron/scheduling side, see how to schedule Cryptohopper MCP workflows. For the Telegram/email output side, see how to send reports to Telegram, Discord, or email.

Pattern 6 — The constrained question

Sometimes you know the agent will over-fetch or over-reason if you let it. Constrain it explicitly.

Use only ticker data for this. Don't pull candles or orderbooks.
Across the top 20 Binance pairs, show me which have spread wider
than 10bp right now.

Answer using at most 3 MCP calls. Pull ETH/USDT ticker from Binance,
Kraken, and Coinbase. Tell me the cheapest venue.

Keep the lookback to 100 candles or fewer. Compute RSI(14) on 1h
for BTC/USDT on Binance and give me a one-sentence read.

Useful when you're running low on quota, prototyping a prompt, or deliberately staying inside a cost envelope. The agent will honour the constraint.

Pattern 7 — The compare-and-explain

Shape: "Compare [A] and [B] on [axis]. Tell me which is [stronger / better / more liquid]."

Compare the 4h trend for ETH/USDT and BTC/USDT on Binance. Is ETH
showing relative strength or weakness vs. BTC over the last 48 hours?

Compare volume profiles for SOL/USDT and AVAX/USDT over the last
7 days. Which has been more active?

Compare the Binance and Coinbase orderbooks for BTC/USDT right now.
Which has the deeper book within 1% of mid? Is the depth balanced
on both sides for each venue?

Useful for relative-strength analysis, venue comparison, and spotting divergences between correlated assets. The agent does the comparison in one reasoning pass — you don't need to chain multiple prompts.

Pattern 8 — The news-plus-context

Shape: "Here's some news. Identify mentioned tokens, pull live context for each, and tell me what the market is doing about it."

Here's an article:
[paste article]
Identify any crypto tokens mentioned by ticker. For each one:
- Pull the current ticker from Binance.
- Pull 4h candles (last 100 periods).
- Tell me whether the market is reacting to this news, ignoring it,
  or already fully priced it in.
At the end, tell me which (if any) tokens look worth a closer look.

The full discussion of this pattern — what to feed it, what to expect — is in news-driven crypto research with AI. It's one of the highest-value prompt patterns the MCP enables.

Pattern 9 — The self-check

A pattern that sounds small but catches real mistakes.

Before you trust your answer, double-check by pulling the ticker
from a second exchange and confirming price is within a reasonable
range. If they diverge by more than 50bp, tell me and don't commit
to the interpretation.

Good for decisions that matter. Bad data — a stale cache, a thin book, a wonky exchange feed — shows up as a divergence. Asking the agent to sanity-check against a second source catches this about 90% of the time without burning real quota (one extra ticker call).

Pattern 10 — The meta-prompt

Ask the agent about the MCP itself.

What MCP tools do you have available for Cryptohopper market data?
Describe what each one is for.

How much of my Cryptohopper MCP quota have I used this week?

Which exchanges does my Cryptohopper MCP tier give me access to?

Boring but useful. The agent will read the schema directly and reply accurately. Good sanity check if you're not sure something is connected, or if you want to remind yourself what's available.

For the quota mechanics specifically, see how to check your Cryptohopper MCP usage and limits.

Building your own library

The prompts above are starting points. The real win is building your own library — a handful of prompts that fit the way you actually think about markets.

A few suggestions for how to curate it:

• Save prompts that worked. The moment a prompt produced output you'd actually use, save it. Don't trust memory.
• Version them. When you change one, keep the old version. You'll sometimes find the simpler earlier version was better.
• Tag by cost. A tag like cheap (ticker only), medium (ticker + candles), expensive (includes orderbooks) helps you pick the right one for the situation.
• Share with yourself. Paste them into a Notion page, a markdown file, wherever you'll find them. A prompt library you can't find is no library at all.

The meta-principle

If there's one rule worth taking away: be specific, but don't route. Name the pair, the exchange, the timeframe, the lookback, the format you want back. Don't name the MCP tool. The specificity narrows the answer; the absence of tool-naming lets the agent choose the shortest path to it.

That balance — explicit about what, implicit about how — is where most of the good prompts live.