Trading Prompts That Work: How to Talk to an AI Agent
The anatomy of prompts an AI agent can execute correctly, three before-and-after rewrites, and a library of ten templates covering DCA, breakouts, mean reversion, and news.
An AI trading agent is exactly as good as the instruction you give it. That's uncomfortable, because most trading ideas live in our heads as feelings ("buy the dip," "ride the breakout") and an agent cannot execute a feeling. Trading prompts are where intent becomes machine behavior, and the difference between a vague one and a precise one is the difference between an automation you trust and one you disable after a week. This article breaks down the anatomy of an executable prompt, rewrites three bad ones, and gives you ten templates to adapt to your own strategy.
Why most trading prompts fail
The failure is almost never the AI misunderstanding English. It's the English not containing a decision.
Take "buy when it looks oversold." Oversold measured by what: RSI, distance from a moving average, a drawdown percentage? On which timeframe? Buy how much? And what happens if the asset keeps falling, which is precisely what oversold assets like to do? That single sentence leaves four decisions unmade, and whatever fills those gaps (a default, an assumption, a guess) is now trading your money.
Good platforms defend against this by translating your sentence into explicit conditions and showing them back to you before anything runs. That restatement loop is the safety net, and the broader shift toward conversational strategy-building is covered in our piece on natural language trading. But the net works best when there's little to catch. A prompt that already contains every decision gets translated faithfully; a vague one gets translated plausibly. Plausible is not the standard you want for order flow.
The six parts of trading prompts that work
Every executable trading instruction answers six questions. Write them in any order and any phrasing; just make sure they're all answered or deliberately omitted.
- Asset: what exactly to trade. "Tech stocks" is a theme; "QQQ" is an instruction.
- Condition: the observable trigger. A price level, an indicator threshold, a crossover, a news event. If a human couldn't verify it from a chart or a feed, an agent can't either.
- Action: what to do when the condition fires: buy, sell, alert you, tighten a stop, pause another agent.
- Size: the amount per action and the ceiling. "$200 per buy, max 3 open positions" removes two of the most expensive ambiguities in trading.
- Invalidation: the condition under which the agent stops itself, like a per-position stop, a total drawdown cap, or an expiry date.
- Schedule: when the rule is alive: always-on, weekdays only, paused around specific events.
Not every prompt needs all six spelled out. A weekly DCA instruction has no market condition, and that's fine, as long as you skipped it on purpose. And once you start combining several conditions in one prompt (price and trend and news), you're building compound logic with its own trade-offs, which we unpack in multi-condition automations.
Before and after: three rewrites
Rewrite 1: the mood. Before: "Buy SOL when it looks oversold." After: "Buy $100 of SOL when RSI(14) on the daily closes below 30, at most 3 buys per week. Sell each lot at +10% or after 21 days. Pause this agent entirely if it's down 15% overall." The feeling became a threshold, a budget, an exit, and a kill condition.
Rewrite 2: the chase. Before: "Get me into the next big breakout." After: "Watch my three watchlist stocks. When one closes above its 60-day high on volume at least 1.5 times its 20-day average, alert me with the numbers. If I approve within 2 hours, buy $1,000 with a stop 4% below entry." Note the action here is an alert plus approval, not a blind buy. Prompts can delegate watching while keeping the trigger pull.
Rewrite 3: the worry. Before: "Protect me if the market crashes." After: "If my portfolio drops 8% from its 30-day high, cancel all open buy orders, notify me on Telegram, and pause all my agents until I re-enable them." Vague fear became a circuit breaker with a defined threshold and a defined blast radius.
The pattern across all three: nothing clever was added. Decisions that were always going to be made got made by you, in advance, instead of by ambiguity at 3am.
A prompt library: ten templates to adapt
These are written the way you'd type them into Obside's copilot, with brackets marking what you should change. Whatever platform you use, read the restated conditions it hands back before approving; that echo is your last check against ambiguity. Backtest and paper trade anything here before real money touches it.
Dollar-cost averaging
1. Scheduled DCA with an overheat skip.
"Buy [$150] of [BTC] every [Monday] at [10:00]. Skip that week's buy if RSI(14) on the daily is above [75]. Resume automatically the next week." The skip clause stops you from mechanically buying blow-off tops; drop it if you want pure DCA.
2. Drawdown-weighted DCA.
"Buy [$200] of [an S&P 500 ETF] on the [1st] of each month. If the price is down [15%] or more from its [90-day] high that day, buy [$400] instead." Adds a rule for what most people only do in hindsight: buying more when prices are marked down. Cap the boosted amount to what your cash flow tolerates.
Breakouts
3. Volume-confirmed breakout.
"When [ASSET] closes above [LEVEL] with volume at least [1.5x] its [20-day] average, buy [$500]. Stop 4% below entry, take profit at [+12%]. One open position at a time." The volume clause filters the quiet drifts above a level that reverse by lunchtime.
4. Breakout retest entry.
"After [ASSET] closes above [LEVEL], don't buy yet. If price returns to within [1%] of [LEVEL] within [10] days and then closes above it again, buy [$400] with a stop [5%] below [LEVEL]." A patience rule, encoded. You'll miss runaway moves and skip most false breaks; decide which failure you'd rather own.
Mean reversion
5. Oversold with a trend filter.
"Buy [$300] of [ASSET] when RSI(14) daily closes below [30] while price is above its [200-day] moving average. Exit when RSI closes above [55] or after [15] trading days. Max [2] concurrent positions." The trend filter is the load-bearing part: it separates dips in an uptrend from knives in a downtrend.
6. Stretch-from-the-mean reversion.
"When [ASSET] closes more than [2] ATR(14) below its [20-day] average price, buy [$250]. Sell when price touches the [20-day] average again. Skip the signal if [CPI or FOMC] is scheduled within the next [24] hours." ATR keeps the stretch measure honest across volatility regimes, and the macro-skip clause avoids catching event-driven moves that mean-revert on their own schedule.
News and events
7. News-severity pause.
"Whenever high-severity negative news hits [ASSET], pause all my buying in it and alert me on [Telegram] with a summary. Resume automatically after [48] hours without new headlines." News as a filter, not a trigger: this template never trades on a headline, it stops you from buying into one.
8. Event reaction, alert-only.
"If [ASSET] moves more than [5%] in either direction within an hour of its scheduled [earnings report], alert me immediately with the direction and the headline. Do not trade automatically." First prints get revised and headlines fake out; the honest automated action around unscheduled chaos is often information, not orders.
Risk and exits
9. Portfolio circuit breaker.
"If my total portfolio value falls [10%] from its [30-day] high, cancel open buy orders across all my agents, pause them, and require my approval to restart." One rule, written on a calm day, that outranks every strategy on a bad one.
10. Trailing exit.
"For every position this agent opens, keep a stop [6%] below the highest close since entry. Never widen a stop." The last clause matters most; widening stops mid-trade is how small losses grow up.
From prompt to running agent
A good prompt is the first third of the job, not the whole job. The pipeline that follows it (restatement and approval, backtest, paper trading, then a small guarded live run) is what turns a well-phrased instruction into an automation with evidence behind it. We walk that full sequence in how to create an AI trading agent.
Two habits will save you more than any template. First, reread the platform's restatement every single time, even for edits; on Obside the copilot re-echoes the conditions after each change precisely because a one-word tweak can move a decimal. Second, keep a written copy of every prompt you run and date it. Vague prompts sit near the top of our list of AI trading mistakes that cost real money, and the ones that hurt most are the ones nobody can reconstruct afterward.
Where to go from here
Steal the templates, but keep the discipline they encode: every bracket you fill is a decision made in daylight instead of delegated to ambiguity. Start with the one closest to something you already do manually, run it on paper, and edit it with the evidence in front of you. When you want to type one of these into a copilot that restates it, backtests it, and runs it behind your risk limits, Obside is built for exactly that loop.
Educational content only. This is not investment advice. Trading involves risk, including possible loss of capital.
FAQ
A trading prompt is a plain-language instruction that an AI trading platform translates into a running automation: monitored conditions, orders, and risk limits. It differs from a chatbot prompt in that the output isn't text but behavior, potentially with money attached. That's why good trading prompts read less like conversation and more like a rule: asset, condition, action, size, invalidation, and schedule, stated explicitly.