By "prompts" I took it you meant something unique.
I am following the usual mechanical backtesting steps. Verify data exists for the dates of the backtest and the identity of the stocks. Norgate provides historical constituents of the S&P500 and Nasdaq100 overtime and their adjusted and unadjusted open, high, low, close and daily volume. After coming up with initial strategies from gtr or algo I submitted tests following the same procedure as at gtr: proposed the sub-universe, run steps each creating a smaller sub-universe based on criteria X based only on price and volume changes as specified (9 month momentum, say, or more complicated). Propose additional steps to further reduce how many stocks Test 1,2,3,5,10,15, 20 stock strategies. Get sharpe ratio, UI, betas vs. Spy and NDX, # of drawdowns over 10%, max dd, time to recover, worst drawdowns, median and mean years, rolling time periods, worst rolling time periods, fragility tests (remove Y number of biggest winners and see result), "mound of toast" type examination (alter criteria to see if it is a coincidental "best" or an entire range is good), examine in known market/economic stress periods of recessions, crashes, covid, etc.....
The usual suspects.

Test holding periods, turn-of-the-month issues, assume .1% spreads and no taxes or inflation for backtest purposes, determine if googlesheets can do a calculation or where you would get ongoing data (stockfetcher, say). Combine strategies to find return/risk mixes that would have been good. Never use today's prices/volume but only last market closes and don't trade until the next market close (I haven't tested many with time delays until purchase except when I did test using any of the last 5 to first 5 trading days of the next month to trade). I tested in and out of made-up time samples.

There is nothing unique in terms of prompt writing in any of this. AI is fully conversant with backtesting issues and statistics. And now it can create googlesheets for you.