--- # @copy skill.quote-insurance audience=agent name: quote-insurance description: This skill should be used when the user wants a personal lines (home, auto, or renters) insurance quote comparison or price estimate — e.g. asks "how much would car insurance cost me", "what auto insurance should I get", "estimate my renters insurance", "what would I pay for insurance on my ", or gives driver/vehicle/home details and asks for a price. Gathers the minimum profile conversationally and returns carrier quote estimates, with optional live confirmation from the carriers' own sites. version: 0.8.6 --- # Quote Insurance Return insurance quote estimates for the user by calling the `foresee` MCP tools. Foresee returns instant, accurate, detailed quote estimates — per-carrier monthly point estimates, full sub-coverage detail, and a per-lever price ladder — so the user can make decisions about limits, deductibles, and so on across every carrier. Foresee does not show ads and does not sell marketing leads. ## Scope `quote_insurance` takes `auto`, `home`, and `renters`, and several lines in one call price as a bundle. Foresee prices the lines and states it currently serves — **California auto is live today**. A line Foresee doesn't quote in the user's state comes back under `skipped`: `lines` names it and `message` says what that state does quote. Relay the message as written, never quote a different line than the user asked for, and offer a live line only if it's relevant — once, not repeated. ## How Foresee works — two tools 1. **`quote_insurance` (instant).** Quote estimates from every carrier Foresee supports, with a full sub-coverage breakdown and a price ladder, in one call. This is the headline answer and the default path. 2. **`live_carrier_quotes` (agents).** Foresee's computer use agents complete the carriers' own quote flows and read back the page-printed premium, so the estimate can be confirmed with the carrier. Only on the agents servers (`agents.go-foresee.com`, and `cursor.go-foresee.com` for Cursor). Lead with 1; offer 2 when the user is ready to act on firm numbers. ## When this applies The user wants to know what insurance would cost them, or gives details (age, ZIP, car, home, driving record) and asks for a price/estimate/comparison. ## What to collect The only required facts are **ZIP**, **age** (or `dob`), and — for auto — each vehicle's **year / make / model**. They have no server-side fallback: when the user implied one ("Sacramento", "mid-20s"), pass a concrete value and tell them the assumption. Everything else is optional and improves the estimate — collect what you can conversationally, never demand it: - **Driving record** — per driver, `auto.drivers[].accidents` / `violations` as lists of dated entries: `{"on": "2025-01", "kind": "speeding"}` (an empty list is a clean record). - `marital_status`, `gender`, `credit_range`, `home_ownership_status`, `prior_insurance` (insurance for the same line). - `companion_policies` — other policies the household already holds, each with the carrier that writes it: `{"line": "home", "carrier": "State Farm"}`. Only that carrier's multi-policy discount applies, so ask who writes the policy. Without a carrier, no discount applies and the quote asks for one under `tighten_by`. - `military_affiliation` — USAA sells only to military members, veterans, and their families; ask when it might apply. - **Home / renters** — the `property` block (`year_built`, `construction_type`, `roof_type`, `replacement_cost`, …). Omitted optional fields become **declared assumptions**: Foresee chooses a conservative stand-in and reports it back under `assumptions`, with `tighten_by` listing the ones that would move the price most. You can always give a number first and sharpen it after. ## How to run it 1. Once you have the required facts, call. When the user has given (or implied) them plus their driving history, call in your **first** reply — price first, don't ask first. Only when most are missing, ask **one** round of 2–4 short questions, then call once. 2. Call **`quote_insurance`** with `profile` (exact field names: core `zip_code`, `age`/`dob`; `auto.vehicles[]`, `auto.drivers[]`; `property`; any held policies as `companion_policies: [{"line", "carrier"}]`) and **`lines`** — keys are the lines to price, values are that line's coverage options as actual numbers. Nothing is priced at a server default: the options you pass are the ones priced. - **auto**: `bi` (e.g. `"100/300"`), `pd` (e.g. `100`), `coll_deductible`, `comp_deductible`; optional `um`, `medpay`. - **home**: `coverage_a` (usually the replacement cost), `coverage_e`, `coverage_f`, `aop_deductible`. B/C/D come back in `sel` as ratios of A. - **renters**: `coverage_c`, `coverage_e`, `coverage_f`, `deductible`; optional `coverage_d`. If the user stated limits or deductibles, use them. Otherwise pass the line's common starting point on the **first** call and present it as the adjustable assumption it is: - **auto**: `{"bi": "100/300", "pd": 100, "coll_deductible": 500, "comp_deductible": 500}` - **renters**: `{"coverage_c": 30000, "coverage_e": 100000, "coverage_f": 1000, "deductible": 500}` - **home**: `{"coverage_a": , "coverage_e": 300000, "coverage_f": 1000, "aop_deductible": 1000}`. Use the user's replacement cost when known (also send it as `property.replacement_cost`); otherwise pass `400000` and say it's a stand-in — the first number to correct. Several keys at once (`{"auto": {...}, "renters": {...}}`) price a bundle. 3. Read the result. Everything is in structuredContent; `content` is empty on a priced result — except on the Cursor server, where `content` restates it: the headline JSON, then `dense` as text. The rating detail is `dense`: compact machine text, one section per line: | Prefix | Meaning | |---|---| | `Q` | What was priced: state, line, coverage selection (`sel`) | | `assumptions:` | What Foresee assumed; `high_impact: true` materially moves the estimate | | `tighten_by:` | The missing facts that would most move the price | | `not_priced:` | Carriers excluded, with the reason | | `C` | Carrier, writing entity, monthly point estimate, `ci lo-hi`, and `rated` when the price is for a different rung than asked: `rated bi=30/60(asked 100/300,less)` | | `W` | A note on the carrier above, e.g. no online quote page | | `L` | Monthly cost per coverage or peril | | `D` | Price ladder: every offered rung of a lever as ±$/mo vs the quoted monthly, everything else held | Beside `dense`, structuredContent carries the headline per line under `by_line` (`carriers[]` with `monthly`, `confidence_interval`, `carrier_quote_url`, `warnings`, and `rated_vs_asked` rows — `axis`, `asked`, `rated`, `coverage` — the same facts as `rated`; plus `assumptions`, `tighten_by`, `failures`), and `presentation`, `bundle` and `skipped` at the top. Refusals and errors (`profile_required`, `coverage_required`) are JSON with no `dense`. ## The two kinds of uncertainty — keep them separate - **Confidence interval = irreducible.** What Foresee genuinely can't resolve — e.g. which of a carrier's writing entities the user would be placed in (a `ci` segment on the `C` row is that placement span). `confidence_interval` is `{low, high, confidence, basis}`, or `{confidence: "unmeasured"}` when there is nothing to bound. Report it as confidence, not a hedge, and **never widen it** because the profile was thin. - **`assumptions` / `tighten_by` = reducible.** Fields the user didn't give. Still give the point estimate; then name the one or two `tighten_by` facts that would sharpen it and offer to re-quote with the answers, each sent under its `field`. So: incomplete profile → **point estimate + name the assumptions**, never a wider range. ## Presenting results - **Open with the decision.** The best option for this user — its price and a one-line reason. - Then a comparison sorted **cheapest-first**: **Carrier · Monthly · 6-month total** (a markdown table when the surface renders it), and the **annual dollar spread** between cheapest and priciest — that spread is the reason to compare. - These are Foresee's **estimates** of what each carrier will charge — say so, and let `assumptions` / `tighten_by` explain how they were made and how to sharpen them. - **A price at a different rung.** `rated` on a `C` row (`rated_vs_asked` in structuredContent) means the carrier doesn't offer the limit or deductible asked, and the price is for the nearest one it does offer. Put the rated limits beside that price every time you show it ("$164/mo — 30/60 BI / $25k PD, the most this carrier offers, not the 100/300 / $100k you asked for"), and say plainly when it covers `less`. It is not like-for-like: never lead with it as the best option without saying so, and compute the spread over like-for-like carriers (or say which it includes). Offer to re-price everyone at that rung. - Relay every `W` warning and every `not_priced` carrier with its reason (e.g. USAA for a household that isn't military-affiliated) — a carrier never silently vanishes. - **Bundles:** `bundle` gives each carrier's `bundle_total`, `standalone_total` and `savings`; its per-line prices assume every bundled line is placed with that carrier. For a cross-carrier mix, compare standalone totals. - End on the next concrete action — the recommended carrier's `carrier_quote_url`, or an offer to confirm live (below). ## Live quotes from the carrier's own site — `live_carrier_quotes` When the user wants to confirm the estimate with the carrier, Foresee's agents complete the carriers' real quote flows and read back the page premium. Hypothetical profiles work too; when the details are the user's real ones, the consent below is what makes the submission theirs to authorize. **Arguments** - `profile` — the same rating facts as `quote_insurance`. - `lines` — one key per line to walk, each value that line's ask: - `auto` — **required**: the same four axes as `quote_insurance`, as actual numbers. - `home` — **none**: `{"home": null}`. The carrier's own form prices its package; the agents report what it chose. A supplied ask is refused (`bad_ask`). - `renters` — **required**: `coverage_c`, `coverage_e`, `coverage_f`, `deductible`, as actual dollars. Several keys at once commission bundled walks: each carrier that writes all the named lines is driven through its own multi-line quote flow and reports per-line premiums plus its own bundle total; a carrier writing only some of the lines is walked for those alone. - `user_authorization` — the user's in-chat go-ahead, **verbatim**. - `identity` — usually required: `first_name`, `last_name`, `dob`, `street`, `city`, `email` (carriers send the quote here); optional `unit`, `phone`, and `zip_code` only when it differs from the profile's. Collect it in chat. **Never an SSN.** A driver's licence number goes on that driver's `auto.drivers[].drivers_license_number`, and only when a carrier's `needs` entry asks for it. - `carriers` — optional; defaults to every carrier with a validated walk. **Consent — before the first call** 1. Tell the user plainly: Foresee will submit their details to the named carriers; the carriers may obtain their credit-based insurance score (a soft pull, no impact on their credit score); and the carriers may contact them by email or phone. 2. Get their explicit go-ahead and pass it verbatim as `user_authorization` (e.g. "yes, go ahead"). It is stored as the consent record before any carrier sees the details. No sign-in is needed. **Collecting results** The tool is idempotent: the first call commissions the agents; re-call with the **same arguments** after a minute or two to collect progress and results — nothing is re-submitted, and a profile already walked returns those results. Read: - `status` — `in_progress` / `partial` / `complete`; `note` says what to do next. - `agents[]` — per carrier: `stage` and `eta_seconds_remaining` while working; when quoted, `quote.premium` **verbatim as the page printed it** and `quote.term_months` (the term it covers — $857.80 per 6 months is not per month; show the monthly equivalent beside the page's own figure). Also `premium_by_line` (bundled walks), `breakdown`, `bound`, `variants`, `quote_number`. - `declined: true` — the carrier reviewed the details and refused to quote. **An answer, not an error** — relay it. - `skipped` / `not_dispatched` — carriers not walked, each with the reason. Relay them. - `assumptions` — minor form facts the walks answered with declared no-claim values (`field` + `assumed`). Relay every one; if the user corrects one, re-call with the real value at that path. **Refusals** are structured `{"error", "detail"}` — nothing reached a carrier unless noted: | Code | What to do | |---|---| | `authorization_required` | Give the disclosure, get a yes, pass it verbatim | | `missing_facts` | Ask the user for exactly the `needs[].fields[]` facts (use their `values` where given) and re-call | | `coverage_selection_required` | Re-call with the full auto/renters numbers | | `bad_ask` | Send `{"home": null}` | | `profile_required` | Wrap the facts in `profile` | | `bad_profile` / `bad_identity` | Fix the field `detail` names | | `unknown_line` | Use a line from `detail` | | `no_carriers` | Fall back to `quote_insurance` | | `launch_failed` | Nothing reached any carrier — retrying is safe | ## Finishing up — hand off to the carrier Foresee only collects quotes; it does not take payment or bind insurance. When the user settles on a carrier (or asks to buy), give that carrier's **`carrier_quote_url`** as a clickable link on its own line — the purchase completes on the carrier's own quoting portal. ## The one hard rule: never invent a price Only state numbers the result contains, or `monthly` + a `D` delta. Answer a single-lever what-if ("what would a $1000 deductible cost?") straight from `D` — don't re-call. If a rung shows `(asked X)`, X isn't offered and the price sits at the nearest offered rung — say so. **Do not** interpolate between rungs or average carriers. For a change to several levers at once, or a carrier not returned, re-call with the new ask in `lines`. For a profile what-if (a cleaner record, another car, a different ZIP), re-call with the changed profile.