Prompting Studio

Studio works best when prompts are specific, bounded, and testable. Treat the AI as a strategy drafting partner: give it the thesis, ask it to draft a strategy, and then use backtesting to decide whether the idea deserves more attention.

Prompt format

A strong prompt usually includes:

  1. market or venue,
  2. signal,
  3. action,
  4. entry threshold,
  5. exit rule,
  6. risk limit,
  7. data source,
  8. deployment intent.

Template:

Build a [venue or market family] strategy that [action] when [signal and threshold].
Use [data source] if available. Cap exposure at [amount]. Avoid [bad conditions].
Exit when [exit condition]. Backtest before deployment.

Examples

Build a Kalshi BTC mean reversion strategy. Use Coinbase BTC spot as edge data. Buy YES only when BTC is up more than 1% over 15 minutes while the prediction market has moved less than 3 cents. Avoid spreads wider than 5 cents, cap exposure at $250, and exit if BTC momentum reverses.
Create a weather strategy using NWS data. Fade YES prices when the market is at least 12 cents above the forecast-implied probability for 20 minutes. Do not trade stale forecasts, do not enter in the final hour, and cap exposure at $150 per market.
Draft a spread capture strategy for a supported Polymarket event. Quote only when the spread is at least 4 cents, keep orders small, cancel stale quotes quickly, and stop quoting if inventory exceeds $200.

Ask for a Structured Draft

When you want a strategy you can inspect, ask for a Studio-ready draft:

Turn this into a Studio-ready strategy and explain each field before running a backtest.

You can also ask Studio to revise a draft:

Keep the same thesis, but lower max exposure to $100, add a stale-data pause, and make exits more conservative. Do not change the market selector.

Ask for a backtest

Backtests are most useful when the prompt frames them as a filter:

Backtest this draft over the longest supported window. I want to know whether the idea survives realistic fills and fees, not just whether it had one good trade.

Useful follow-ups:

  • "Which assumption drives most of the result?"
  • "Did this depend on one outlier market?"
  • "How many trades were simulated?"
  • "How sensitive is it to spread limits?"
  • "What happens if we cut max position in half?"
  • "Show the worst drawdown and explain it."

Ask for deployment conservatively

Live deployment should be explicit:

Deploy this strategy to my runner with the same risk limits used in the backtest. Start small, keep monitoring enabled, and do not increase exposure without asking me.

Avoid prompts like:

Go live and optimize it.

That gives the AI too much room to change behavior without a clear approval point.

Prompting rules for AI agents

If you are an AI agent operating Studio:

  • Make the user's thesis visible in your response.
  • Produce a clear strategy draft when the user asks to build.
  • Use conservative defaults only when the user has not specified a value.
  • Mark assumptions clearly.
  • Backtest before deployment.
  • Never claim that a strategy is safe, guaranteed, or risk-free.
  • Never increase exposure to improve a backtest without explicit approval.
  • Prefer "this idea should be rejected" over overfitting a weak result.

Common failure modes

Too broad

"Trade weather markets" is not enough. Ask which geography, timeframe, forecast data, and action.

No risk cap

A strategy without max exposure is incomplete. Add a cap before testing.

No exit

Every entry needs an exit. Exits can be signal reversal, time to close, take profit, stop loss, stale data, or manual review.

Data mismatch

Do not use edge data that would not have been known at the simulated time. Backtests should not rely on future information.

Performance chasing

If each revision only tries to improve a metric, stop and restate the thesis. A strategy should have a reason to exist outside the backtest.