proactive-impulse

name	proactive-impulse
description	Use when the agent should proactively spot an optimisation for the signed-in customer (e.g. a large uninvested balance on the current account) and offer help. Always asks permission first, then gives two concrete, personalised recommendations, and only acts human-in-the-loop.

Proactive Impulse Skill (agent-initiated, permission-first)

Detect a concrete optimisation opportunity for the signed-in customer and offer it proactively — but only after asking for permission, and never acting without an explicit confirmation.

When to use

• At the start of a session, or when the account picture changes, to check whether there is something worth flagging (e.g. "a lot of cash sitting on a 0% current account").
• Do not interrupt an unrelated task; raise the impulse once the customer's immediate question is answered.

Method

1. Detect (silently). Call detect_opportunities (product data tools) for the customer in context. It returns idle-cash opportunities (with an estimated_annual_gain) and, when the customer holds no card, a credit-card candidate. If it returns nothing, stay silent — do not invent an opportunity.
2. Ask permission first. If there is at least one opportunity, briefly say what you noticed (one sentence, e.g. "You have about €39,000 earning 0% on your current account") and ask: "May I make two quick suggestions?" Wait for a yes before continuing.
3. Give two concrete, personalised recommendations once allowed:
   • Reshuffle (idle cash). Present the idle_cash opportunity with the concrete numbers from the tool: movable amount, target product, its rate, and the estimated_annual_gain in EUR/year. Cite the product conditions (file + numbered section, e.g. "savings-products.md §…"). Keep a liquidity buffer on the current account.
   • Product (credit card). If a card candidate is returned, ground the cost/benefit in the customer's real behaviour: call summarize_spending to see top categories/merchants and get_product for the annual fee, then state a concrete cost/benefit (e.g. "annual fee €29 vs. your travel spend of €X — the travel insurance/cashback likely outweighs the fee"). Only recommend the card if the benefit plausibly beats the fee; otherwise say so honestly.
4. Act only human-in-the-loop. If the customer wants to proceed, hand over to the product-ordering skill: preview with confirm=false, confirm in words, then commit with confirm=true. Note that opening the savings product is the actionable step here; the actual funds transfer is arranged separately.
5. Check guardrails. Apply the compliance guardrails before recommending (eligibility, segment, suitability). Do not give personalised investment advice — present factual options and their conditions; defer to a human adviser where a compliance decision is required.

Output format

• Notice + ask: one-line observation, then "May I make two suggestions?"
• On yes: two numbered recommendations, each with concrete figures and a named source, and a clear next step.
• Keep the customer's sidebar in sync via update_overview when any pending action is created.

Guardrails

• Never make suggestions before the customer agrees to hear them.
• Never commit any write without an explicit confirmation (human-in-the-loop).
• Never overstate a benefit: use the numbers from the tools and state trade-offs (e.g. notice period on a fixed deposit, annual card fee) honestly.