Using your Dust agent as a thinking partner

The most useful habit: ask before assuming

This principle holds regardless of platform: ask your agent directly what it can do, before even phrasing your final request. A Dust agent configured with the Spendesk MCP knows the exact list of available tools. So for a given need, it can tell you:

• Which tools it would use.
• Whether the need is fully covered, or only approximated.
• What workaround exists if coverage is partial.
• How to turn all of that into reusable agent instructions.

What we already know is out of reach for the MCP

Three tools come up repeatedly in client requests but don't exist in the current Spendesk MCP — this limitation is identical on Dust and Claude, since it comes from the MCP server, not the platform:

• get_budgets — no native budget-vs-actual comparison.
• get_payable_events — no event log on a payable. For a supplier invoice, the approval chain (who approved it, when, with what note) can be read with get_invoice_by_id.
• get_bank_transactions — no direct transaction-level bank reconciliation.

And a limitation that recently evolved: write tools are in Beta and off by default. Your agent can read your Spendesk data, and 26 write tools exist — suppliers, payables, purchase orders, chart of accounts, accounting exports, analytical fields — but each write permission must be enabled by an Admin or Account Owner, then ticked by the user. There is still no tool to approve, pay, read budgets, or create or change cards, users or cost centers.

The prompt to use for any new need

Here's my need: [describe your business situation in 2-3 sentences, with context — your role, desired frequency, end goal].

Based on the available Spendesk MCP tools (spendesk_analyze_spend, spendesk_get_due_invoices, get_payables, get_suppliers, get_purchase_orders, get_requests, list_cards, get_cost_centers, get_settlements, get_wallet_summary, etc.):

1. Which MCP tools would you use to answer this need?
2. Is this need fully covered by the Spendesk MCP, or is there a limitation (e.g. no webhooks/real-time alerts, write tools in Beta and off by default, no approval or payment tool, no budgets)?
3. If it's not fully covered, propose a concrete workaround.
4. Draft me reusable agent instructions for this need.

A concrete example

Closing the loop

Once your agent has assessed your need, what comes next depends on its answer:

• Full coverage, one-off need → paste the instructions into a new agent.
• Full coverage, recurring need → add a Trigger.
• Recurring need + team-wide visibility → publish it to Slack.
• Partial coverage → note the proposed workaround, and keep in mind that the skills library keeps growing as the MCP is updated.

For every role: this prompt works just as well for a very specific need as for a broader project — the more precise your initial description, the more immediately actionable the answer will be.