Skip to main content

Operations and Safety Prompts for Trading Agents

These prompts are for the work around the trade. They keep your AI trading agent honest about state: which charts are open, which alerts are armed, which strategies survived the last reconnect, which permission scope is in effect, and which commands are dangerous enough that the agent must stop and ask first.

Every prompt here is built around tools that are operational today. If a capability is not yet supported (timeline scrubbing, chart-render confirmations, etc.), say so in the prompt instead of letting the agent improvise.

Related references: Platform Tools, Workflows and Safety, NinjaTrader 101, and Risk Management Basics.

Pre-session readiness check​

Before the agent does anything else, force it to prove the environment is sound.

You are my CrossTrade MCP trading agent. Before any other action, run a readiness check:

  1. Call GetMcpCapabilities and confirm addon_version is v1.13.0 or higher and backtest_engine.available is true.
  2. Call McpSelfTest with deep:false and confirm every category passes.
  3. Call GetConnections and tell me which NT8 connections are Connected.
  4. Call ListAccounts and tell me which accounts the token can see.

If any step fails, stop and explain exactly what is broken. Do not place orders, deploy strategies, or write NinjaScript until the readiness check passes.

This is the single most useful prompt to put at the top of every session. The MCP server does not invent state. If the add-on is offline or the wrong version, the agent should fail loudly, not silently retry.

Open a chart for visual context​

Charts are read-only context for the human reviewer. The agent does not need a chart to backtest or deploy, but you usually want one open so you can watch what the agent is reasoning about.

Open a 5-minute MES chart for the current front month so I can watch while you work. After OpenChart returns, confirm whether the chart was newly opened or already_open:true, and then list the chart's identifying fields so I can verify in NT8.

Do not claim a drawing is visible on the chart unless the tool result explicitly reports a chart render confirmation. AddDrawing registers an intent; the visible render is a separate concern.

This keeps the agent honest about what the tool guarantees. OpenChart does open a window; AddDrawing registers a drawing intent the addon tracks.

Alert lifecycle: arm, verify, clean up​

Alerts are how an autonomous agent tells you something matters without you watching the screen. Make every alert have a known lifecycle.

Create a log alert for MES that fires when price crosses above the prior day's high. Then:

  1. Call ListAlerts and confirm the new alert appears with the ID you expected.
  2. Tell me the alert ID, the trigger condition, and the destination channel.
  3. At the end of this session, remind me to RemoveAlert by ID so we don't leave orphaned alerts behind.

This is a safe CRUD pattern: create, list, identify, and plan removal. The agent should never silently leave alerts running.

Risk notification to your phone​

EmitMcpAlert routes alerts through your configured destinations (in-app, email, Discord, Telegram). The agent can use it to call you out of a meeting if something matters.

Send me an MCP alert with subtype risk_guard, title "Risk guard armed", and message "Sim101 is flat and daily lockout monitoring is active." Use account Sim101 and deduplicate for 60 seconds. Confirm the alert payload you sent and remind me that external delivery depends on which channels I've enabled in alert preferences.

This uses the alert relay honestly without promising email, Discord, or Telegram delivery unless those destinations are configured.

Workspace hygiene​

Workspaces define the user's NT8 layout. An agent that overwrites a workspace can destroy your screen real estate. Always confirm before mutating.

List my NT8 workspaces using ListWorkspaces. Tell me which one is currently loaded.

If a workspace named Morning_Risk_Check exists, describe what's in it. If it does not exist, ask me before calling SaveWorkspace under that name. Never call LoadWorkspace or overwrite a workspace without my explicit confirmation in this message.

Default rule: workspace reads are safe; writes (save, load) need explicit user approval per call.

Reconnect recovery verification​

NT8 connections drop. CrossTrade's reconnect-restore flow attempts to bring deployed strategies back. The agent's job after a reconnect is to verify what survived, not to assume the recovery worked.

I just reconnected NT8. Verify recovery state across the system:

  1. GetConnections (everything that should be up is up).
  2. ListDeployedStrategies (which deployments are in the registry).
  3. For each deployment, GetDeployedStrategyState (confirm is_trading and live block).
  4. GetAllPositions and GetAllOrders across every account.
  5. GetWatermarks for accounts where I track daily P&L.

Report anything that is faulted, disabled, or in an unexpected state. Do not auto-restart anything. Tell me what you would do and wait for my confirmation.

This prompt does not assume the reconnect policy worked. It asks for evidence.

Permission scope smoke test​

Most users start with both mcp:read and mcp:trade tokens. Verify which one the active client is actually using before the agent tries to do anything destructive.

I want to confirm which MCP scope this client is using. On Sim101, attempt to place a tiny limit order ten ticks below the current bid for one MES contract. If the order would be rejected for scope reasons, report the exact error from PlaceOrder. If the order is accepted, immediately cancel it and tell me the order ID round trip.

Do not switch tokens or attempt to bypass the permission model. The correct outcome on a read-only token is denial.

Useful when setting up a new AI client and you're not sure which token you authorized.

Dangerous-command gate​

Short prompts are dangerous prompts. Make the agent show its work before mass actions.

Cancel all orders.

Use this instead:

Show all working orders across all accounts. Group them by account and instrument with order IDs. If you would recommend canceling any of them, propose a per-order cancel plan first. Do not use account-wide cancel tools (CancelAllOrders, FlattenEverything) until I confirm the exact account list and instrument list in a follow-up message.

For a real emergency, be explicit about scope:

Emergency on Sim101 only: flatten all ES and MES positions and cancel working ES and MES orders. Before acting, repeat the account name and instrument list back to me word for word. Do not touch any other account or instrument.

End-of-session shutdown​

Closing the trading day cleanly is its own discipline. An agent that "forgets" to flatten before the bell is a real loss.

Run my end-of-session shutdown on Sim101:

  1. ListPositions and ListOrders for Sim101. Show me everything that is still open or working.
  2. For each deployed strategy I want to retire today, propose a StopStrategy call by deployment_id. Wait for my confirmation per strategy.
  3. For working orders that are not protective brackets on an open position, propose a cancel plan. Wait for my confirmation.
  4. After my confirmations, execute, then re-read positions and orders and confirm the account is flat.
  5. RemoveAlert for any alerts you created this session so we don't carry orphans into tomorrow.

Conclude with an EmitMcpAlert summarizing the shutdown so I get a notification on my phone.

This is the kind of multi-step ritual MCP was designed for. Without an agent driving it, you do these one click at a time in NT8.

Agent handoff between sessions​

If you start the day on one MCP client (say Claude Desktop) and continue in another (say Cursor), the new agent has zero context. Force it to rebuild context before it acts.

I'm continuing a session that started in another MCP client. Before you do anything, rebuild context:

  1. ListDeployedStrategies and tell me what is registered as deployed.
  2. GetDeployedStrategyState for each one, focusing on is_trading and any faulted flag.
  3. GetAllPositions and GetAllOrders across accounts I trade.
  4. ListAlerts and ListMcpJobs (status=running) so I know what is in flight.

Summarize the state of the session and the open obligations before suggesting any new action.

Back to MCP Trading with AI.