A useful insurance prompt has five parts in a specific order - role, context, artifact, constraint, format - and a sixth, the reason-code clause, that satisfies NAIC §4 and survives a market-conduct exam. The Dallas commercial-property underwriter, the Atlanta auto-and-GL adjuster, the Hartford renewal producer, and the chief actuary's pricing analyst each need a different role framing, a different context payload, a different artifact target, a different constraint set, and a different output format - but they each need the same reason-code discipline at the close. This lesson walks all four side by side: the UW triage prompt that turns ACORD 125 + 140 + SOV + 5-year loss run into a Cytora-style triage memo; the claims summary prompt that turns a Hi Marley SMS thread or a Five Sigma recorded-statement transcript into a clean FNOL; the renewal-narrative prompt that turns last year's policy + the operations update + the loss-prevention investment record into a stewardship narrative for a Travelers / Chubb / Hartford / CNA submission; and the rate-filing-memo prompt that turns an Akur8 model output into a SERFF-ready actuarial memorandum citing ASOP 56, ASOP 41, and ASOP 23. Each prompt ends with the reason-code clause: "List the variables driving each adverse-decision recommendation. Flag any variable that could act as a protected-class proxy. Cite the policy form by edition and the controlling authority by citation. If you are uncertain about a fact, mark it [verify] rather than invent."

The Five-Line Preamble Every Insurance Prompt Needs

Before role, context, artifact, constraint, and format, every insurance prompt opens with a five-line preamble that constrains the entire conversation. The preamble is identical across UW, claims, producer, and actuarial work - only the role line below changes. Memorize it.

Line 1 - Role. "You are a [licensed underwriter in the {LOB} line at a {carrier-type} carrier writing in {state list}] / [licensed adjuster handling {LOB} claims at a {carrier-type} carrier in {state list}] / [licensed producer at a {agency-type} agency placing {LOB} in {state list}] / [credentialed actuary supporting {LOB} rate filings in {state list}]." The role line names the license posture and the jurisdictional surface; it does not name a persona ("expert underwriter") because persona language drifts the model into rhetoric instead of constrained output.

Line 2 - Jurisdiction. "I write in {state list}. NAIC Model Bulletin §4 reason-chain discipline applies. State overlays in effect: {Colorado Reg 10-1-1 / NY DFS Circular Letter 2024-7 / Connecticut MC-25-8 / Nevada Bulletin 24-006 / Texas TDI / Florida OIR} as relevant." The jurisdiction line forces the model into the regulatory frame; without it the model defaults to a generic posture that produces non-compliant artifacts.

Line 3 - Form edition. "Cite policy forms by exact edition (e.g., ISO CG 00 01 04 13 not 'CGL form'). If the edition is not given in the source material, mark the citation [verify against dec page]." The form-edition line is the single highest-leverage clause against the wrong-edition failure mode the Five Sigma example produced at L1.

Line 4 - Reason-code policy. "Every adverse-decision recommendation (decline, knockout, quote-with-restriction, denial, ROR, SIU referral) must enumerate the variables driving the decision and flag any variable that could act as a protected-class proxy. If the recommendation rests on an AI score (Shift, CCC, Federato, Cytora, Akur8, Earnix), state the score as a signal and list human-reviewable facts separately." The reason-code clause is what makes the output §4-compliant; missing it produces an artifact the carrier cannot defend.

Line 5 - Uncertainty handling. "If you are uncertain about a fact, mark it [verify] and state what document or system would confirm. Do not invent ISO form numbers, ISO endorsement numbers, NCCI class codes, state DOI circular numbers, case-law citations, or carrier appetite specifics." The uncertainty clause is the hallucination guardrail; the named items (ISO forms, NCCI codes, case law, circulars) are the documented hallucination surfaces L2 Lesson 3 will diagnose in detail.

Role, Context, Artifact, Constraint, Format - The Five-Part Body

The five-line preamble sets the regulatory frame. The five-part body is what produces useful insurance output: role, context, artifact, constraint, format. Each part has a specific structure.

Role. One sentence on what the LLM is doing and who it is doing it for. "Triage this commercial-property submission for the Dallas underwriting desk." Not "be a great underwriter." Not "help me with this submission." The role sentence names the action and the audience.

Context. The named facts the LLM needs and only the facts the LLM needs. The submission packet (ACORD 125 + 140 + SOV) is context. The five-year loss run is context. The carrier's appetite guide for habitational frame is context. The producer's marketing pack is not context unless the prompt action requires it. Context bloat is the second-largest source of bad LLM output (after missing preamble); over-stuffing context drives the model toward verbose paraphrase rather than constrained analysis.

Artifact. The named insurance document the output is going to become. "Triage memo for the Cytora workbench" / "FNOL summary in the carrier's intake template" / "Renewal stewardship narrative for the marketing pack" / "SERFF-ready rate-filing memorandum." Artifact naming is what tells the model which template, which sections, which length, and which voice to produce.

Constraint. The set of things the model must do and must not do. Must do: cite policy forms by edition, enumerate reason codes on adverse decisions, separate AI signals from human-reviewable facts, mark uncertain facts [verify]. Must not do: invent ISO form numbers, fabricate case-law citations, draft coverage opinions without citations, write narrative around protected-class proxies. Constraints are where insurance-specific discipline lives.

Format. The structural shape of the output. Headings, tables, bullet lists, prose blocks, line lengths, JSON schema for downstream PAS/AMS ingestion. Format naming is what makes the output usable by the next workflow step rather than the immediate reader.

Prompt 1 - The UW Triage Prompt for Dallas Commercial Property

Walk to the Dallas 8:14 a.m. underwriting desk. The submission is a 47-building commercial-property packet from a regional retail broker. The chief underwriter has flagged three buildings as habitational frame appetite exception. The treaty carries a $25M single-risk limit. Cat aggregate is 78% consumed for the quarter. The underwriter's prompt needs to produce a Cytora-style triage memo against the carrier's appetite guide, with knockout reasons and missing-info requests in the format that drops into Federato.

System message. "You are a licensed P&C underwriter in the commercial-property line at a regional admitted carrier writing in TX, OK, AR, LA, MS, NM. NAIC Model Bulletin §4 reason-chain discipline applies. Cite policy forms by exact edition. Every adverse-decision recommendation must enumerate variables driving the decision and flag protected-class proxies. If uncertain, mark [verify] and state the document that would confirm. Do not invent ISO forms, ISO endorsements, or carrier appetite specifics."

User message. "Triage this commercial-property submission for the Dallas UW desk. Context: 47-building SOV (attached), ACORD 125 with named insured Acme Warehousing Inc., ACORD 140 with property values and COPE per location, 5-year loss run with two non-cat fire losses ($142K and $87K incurred) and four wind-related losses ($14K to $61K each). Carrier appetite guide attached: declines habitational frame in coastal counties; quotes class 5 light non-combustible with restrictions; binds class 6 fire-resistive standard. Treaty: $25M single-risk facultative cession threshold; treaty cat aggregate 78% consumed Q1. Three buildings flagged 'habitational frame' by the producer - verify against ISO construction codes. Artifact: Cytora-style triage memo with (1) overall appetite recommendation (bind/refer/decline/quote-with-restriction), (2) per-building appetite map with knockout reasons where applicable, (3) treaty-cession recommendation per building over $25M TIV, (4) missing-information requests, (5) reason-code summary citing the variables driving each adverse recommendation and flagging any protected-class proxies, (6) verification flags on the three habitational-frame buildings. Constraint: do not cite ISO form numbers without the edition; do not fabricate appetite specifics; if construction class is ambiguous, request the COPE-specific worksheet rather than guess. Format: structured headings (Overall / Per-Building / Treaty / Missing-Info / Reason Codes / Verification), bullet lists under each, length 600-900 words."

What the prompt does right. Names the carrier-type and state list; loads the artifact-specific context (SOV, ACORD 125, ACORD 140, loss run, appetite guide, treaty) without bloat; names Cytora as the workbench template; demands reason codes with protected-class flagging; constrains hallucination (no invented ISO codes, no fabricated appetite specifics); structures the output for Federato ingestion.

Prompt 2 - The Claims FNOL Prompt for the Atlanta Adjuster

Walk to the 9:00 a.m. Atlanta auto-and-GL adjuster's diary. The water-damage file's Hi Marley SMS thread carries 47 messages over 11 days. The insured stopped responding on day 7. The adjuster needs an FNOL summary that captures the loss facts, the coverage triggers, the parties, the prior-loss check, the first-touch contact plan, and the cycle-time alert - in the format that drops into the carrier's claims-intake template.

System message. "You are a licensed P&C claims adjuster at a regional carrier handling first-party property claims in GA, AL, SC, NC, TN, FL. NAIC Model Bulletin §4 reason-chain discipline applies. Cite policy forms by exact edition; if the dec page form edition is not in context, mark [verify against dec page]. Do not invent ISO endorsement numbers. Enumerate the human-reviewable facts on any AI-influenced decision separately from the model score. State Unfair Claims Settlement Practices Act regimes apply to the FNOL handling."

User message. "Draft the FNOL summary on this water-damage claim for the carrier's intake template. Context: Hi Marley SMS thread attached (47 messages over 11 days, last insured response day 7). Insured policy HO 00 05 [verify edition] with declared dwelling value $384K, separate-structures sublimit 10%, ordinance-or-law endorsement [verify presence], water-backup endorsement [verify presence]. Loss reported day 1 as 'water in basement from sump pump failure overnight.' ISO ClaimSearch prior-loss check returned one 2022 water claim ($14K paid) at the same address. Artifact: FNOL summary in the carrier intake template with (1) date/time/location of loss, (2) reported cause and damages, (3) parties (insured, contractor mentioned in SMS thread), (4) coverage triggers and form-edition citations marked [verify against dec page] where appropriate, (5) ISO ClaimSearch prior-loss handling and the impact on this claim, (6) first-touch contact plan within the 24-hour SLA - given day-7 last response, propose outreach cadence, (7) cycle-time alert flag for the 11-day age. Constraint: do not invent endorsement numbers; if water-backup or ordinance-or-law endorsement presence is uncertain, mark [verify against dec page]. Do not draft coverage opinion - that is a separate artifact. Format: headings per intake template (Loss Facts / Coverage / Parties / Prior Loss / First Touch / Alerts), bullet lists under each, length 400-650 words."

What the prompt does right. Names the state UCSPA regime; loads the Hi Marley thread + dec-page-pending facts as context; marks uncertain endorsement presence [verify against dec page]; integrates ISO ClaimSearch as the deterministic reference; demands the first-touch SLA explicitly; restricts the model from drafting the coverage opinion (a separate artifact in the L2 chapter); structures the output for the intake template.

Prompt 3 - The Renewal Narrative Prompt for the Hartford Producer

Walk to the Hartford-based commercial-lines producer's desk. The renewal account is a $4.2M revenue distribution business with three locations, a 14-vehicle commercial auto fleet, and a 22-employee workers' comp exposure. The incumbent carrier is Travelers; the producer is marketing to Travelers, Chubb, Hartford, and CNA for the renewal. The expiring policy is 60 days out. The producer needs a stewardship narrative for the marketing pack, plus carrier-specific cover emails that name each carrier's appetite hooks.

System message. "You are a licensed commercial-lines producer at an independent agency placing middle-market property/GL/auto/WC packages in CT, MA, NY, NJ, PA, RI. NAIC Model Bulletin §4.2 third-party AI considerations apply to any AI-overlay on agency work. State agent-licensing obligations attach to the producer's signature. Do not invent carrier appetite specifics; if a carrier's appetite hook is unclear, mark [verify with carrier underwriter] rather than guess. Do not fabricate Travelers / Chubb / Hartford / CNA underwriting guidelines."

User message. "Draft the renewal stewardship narrative and four carrier-specific cover emails for the Acme Distribution Inc. renewal. Context: $4.2M revenue distribution business, three locations (Hartford CT, Springfield MA, Worcester MA), 14-vehicle commercial auto fleet, 22 employees, EMR 0.88, five-year loss run with one GL slip-and-fall settled at $42K (2023) and one auto comprehensive at $11K (2024). Incumbent Travelers expiring policy package: $1M/$2M GL, $1M auto CSL, statutory WC, $500K BI, $4.2M total property values. Operations update: opened the Worcester MA location 18 months ago; implemented forklift-certification training program 2024 (training records on file); added two HazMat-endorsed drivers for cosmetics-distribution route (clean MVRs); installed fleet telematics (Samsara) 2024. Loss-prevention investment record attached. Artifact: (1) renewal stewardship narrative for the marketing pack (450-700 words), naming the operations update, loss-prevention investments, EMR trajectory, and risk-profile improvement story; (2) one-page exec summary for the client CFO; (3) four carrier-specific cover emails - Travelers (incumbent retention), Chubb (mid-market multinational appetite), Hartford (small-mid commercial sweet spot), CNA (industry-specialization play) - each naming the carrier's documented appetite hooks for distribution-class businesses. Constraint: do not invent carrier appetite specifics; if a hook is uncertain, mark [verify with carrier underwriter]. Do not draft coverage opinion language. Format: stewardship narrative as prose with headings; exec summary as bulleted one-page; cover emails as standalone email-format blocks with subject lines."

What the prompt does right. Names the producer's state list and license posture; loads the operations update, loss-prevention investments, and expiring-policy specifics as context; names the four target carriers explicitly; demands appetite specifics be verified [with carrier underwriter] rather than invented; restricts the model from drafting coverage opinions (E&O exposure); structures the output as three artifacts in one prompt for efficiency.

Prompt 4 - The Rate-Filing Memo Prompt for the Chief Actuary's Pricing Analyst

Walk to the chief actuary's pricing analyst desk. The Q2 rate filing on commercial property goes to SERFF in three weeks. The Akur8 commercial-property GLM has been rebuilt with the January 2026 model refresh; the SHAP global feature importance and the PDP marginal-effects plots are exported; the bias-test exhibit by protected class is complete; the variable-attribution narrative needs drafting; the actuarial certification under ASOP 56, ASOP 41, and ASOP 23 needs language tying the AI involvement to the actuarial standards.

System message. "You are a credentialed pricing actuary (ACAS/FCAS or MAAA) supporting commercial-property rate filings at a regional admitted carrier writing in TX, OK, AR, LA, MS, NM. ASOP 56 (Modeling), ASOP 41 (Communications), and ASOP 23 (Data Quality) apply. NAIC Model Bulletin §4.3 (testing and validation) and §4.4 (documentation) apply. NY DFS Circular Letter 2024-7's proxy test applies on the NY portion of the book (the carrier does not write NY currently - note this in the memo). Colorado Reg 10-1-1's algorithm-inventory entry applies. Do not invent SHAP values, PDP curves, bias-test ratios, or actuarial methodology specifics. If a value is not in the attached model output, mark [verify against Akur8 export]."

User message. "Draft the SERFF-ready rate-filing memorandum for the Q2 2026 commercial-property base-rate revision. Context: Akur8 commercial-property GLM January 2026 refresh (model output attached: SHAP global feature importance with top 12 variables, PDP marginal-effects plots per top variable, segment-level loss-ratio actual vs. expected, champion-challenger AUC comparison vs. legacy SAS GLM). Bias-test exhibit attached: disparate-impact ratio across protected classes (race/ethnicity by census tract proxy, age band, gender) with no ratio falling below 0.85. Akur8 Discover competitive-filing analysis attached for benchmarking. Proposed overall rate impact: +4.2% statewide TX with +2.8% to +6.1% segment variation. Artifact: SERFF-ready rate-filing memorandum with (1) cover transmittal letter to TDI, (2) actuarial memorandum citing ASOP 56 / ASOP 41 / ASOP 23 and naming the AI involvement explicitly, (3) variable-selection narrative covering the top 12 variables with rationale and protected-class proxy assessment, (4) bias-testing exhibit summary linking to the underlying data, (5) rate-change exhibit by class/territory, (6) NAIC §4.3 testing-validation summary and §4.4 documentation completeness summary, (7) competitive-filing benchmarking summary from Akur8 Discover (benchmark only, not adverse-action driver). Constraint: do not invent SHAP values or bias-test ratios; do not fabricate ASOP language - quote exact standards by section. Do not cite competitor filings as adverse-action drivers. Format: SERFF transmittal format with standard sections; variable-selection narrative as prose; bias-test as table; rate-change exhibit as table; length 2,500-4,000 words across the memorandum."

What the prompt does right. Names the actuarial credential and the ASOPs by number; loads the Akur8 outputs as context; flags Colorado Reg 10-1-1 + NY DFS overlay considerations; demands quote-exact ASOP language; restricts SHAP/PDP fabrication; restricts Akur8 Discover from being cited as adverse-action driver; structures the output as a SERFF-ready package.

Why the Four Roles Need Different Role Framings

The four prompts share the five-line preamble, the five-part body structure, and the reason-code clause - but the role framing differs in ways that matter to the output. The underwriter's role line names a licensed P&C underwriter in a specific LOB; the model produces appetite-disciplined analysis, declines-with-reason-codes, and treaty-cession recommendations. The adjuster's role line names a licensed claims adjuster in specific states; the model produces FNOL handling discipline, coverage-trigger awareness, ISO ClaimSearch integration, and SLA-aware first-touch planning. The producer's role line names a licensed commercial-lines producer at an independent agency; the model produces marketing-pack narratives, appetite-hook specifics, and stewardship language - with E&O guardrails on coverage opinions. The actuary's role line names a credentialed actuary with ASOP obligations; the model produces SERFF-ready memoranda, bias-test exhibits, and competitive-benchmarking analysis with ASOP language quoted exactly.

The cost of getting the role framing wrong is large. A claims-adjuster role framing on a UW triage prompt produces output that focuses on loss-handling rather than appetite. A producer role framing on a claims FNOL prompt produces output that drifts into client-relationship language rather than intake template discipline. An actuary role framing on a renewal-narrative prompt produces output that loads ASOP language into a marketing pack the broker can't use. The role line is not stylistic - it is the model's anchor for which discipline applies to the output.

The Reason-Code Clause That Closes Every Prompt

Every adverse-decision recommendation in any of the four artifacts - UW decline, knockout, quote-with-restriction, claims denial, ROR, SIU referral, rate-impact adverse-action, producer-side coverage gap - needs a reason-code clause that satisfies NAIC §4. The clause has three components: (1) enumerate the variables driving the decision; (2) flag any variable that could act as a protected-class proxy; (3) state the AI signal as a signal and list human-reviewable facts separately. Prompt-level integration is the rule - every prompt in this lesson includes a sentence in the Constraint section that mandates the clause on adverse outputs.

The clause's structural role is that it converts opaque output into auditable output. A Cytora triage memo with "decline - habitational frame in coastal county" is opaque; the same memo with "decline. Variables: ISO construction class 1 (frame), Tier 1 wind-exposure zone (coastal Brazoria County), prior-loss frequency above appetite for class 1 frame habitational. Protected-class proxy assessment: census-tract proximity flagged; verified against current carrier guidance that geographic underwriting in Brazoria County is not a protected-class proxy under TDI Bulletin B-2024-3" is auditable. The audit-defensible version is what survives the §4 exam, the bad-faith complaint, and the producer pushback on the decline.

When to Keep the Prompt and When to Throw It Out

The four prompts in this lesson are templates, not formulae. The underwriter customizes the appetite guide; the adjuster customizes the intake template; the producer customizes the carrier list; the actuary customizes the ASOP citation. But the five-line preamble, the five-part body structure, and the reason-code clause are the durable architecture - they do not change between submissions, between claims, between renewals, between rate filings. The L4 prompt-log retention artifact captures every prompt run against every artifact; the L3 RAG architecture loads dec page, ISO form library, appetite guide, claims-handling manual, ASOP text, and case-law corpus as retrieval sources; the L2 verification discipline catches output failures before the artifact lands on the desk of the next reviewer.

The prompt is the entry point. The reason-code clause is the exit. Between them sits the work the next four chapters of L2 walk through: triage memos, appetite memos, FNOL summaries, coverage analyses, ROR letters, EUO outlines, SIU referrals, renewal narratives, BOR letters, rate-filing memos, treaty-cession recommendations. Each artifact has its own prompt template; each template starts with the same preamble; each template ends with the same reason-code clause.

Key Takeaways

• The five-line preamble (Role, Jurisdiction, Form edition, Reason-code policy, Uncertainty handling) opens every insurance prompt. Without it the model defaults to a generic posture that produces non-compliant artifacts. The preamble is identical across UW, claims, producer, and actuarial work; only the role line changes.
• The five-part body - Role, Context, Artifact, Constraint, Format - produces useful insurance output. Role names the action and audience; context loads named facts without bloat; artifact names the named insurance document the output becomes; constraint sets the must-do and must-not-do list; format specifies the structural shape for downstream PAS/AMS ingestion.
• The UW triage prompt names the Cytora workbench artifact, loads ACORD 125 + 140 + SOV + loss run + appetite guide + treaty constraints, demands per-building appetite mapping with reason codes. Output structures for Federato ingestion. The Dallas 47-building submission with three habitational-frame buildings on appetite exception is the canonical example.
• The claims FNOL prompt names the carrier intake template, loads the Hi Marley SMS thread + dec page + ISO ClaimSearch prior-loss + first-touch SLA. Marks endorsement presence [verify against dec page] rather than invent; restricts the model from drafting the coverage opinion (a separate artifact). The Atlanta water-damage 11-day non-response file is the canonical example.
• The renewal narrative prompt names the four target carriers (Travelers, Chubb, Hartford, CNA) and demands appetite specifics be marked [verify with carrier underwriter] rather than invented. Restricts the model from drafting coverage opinions due to E&O exposure. The $4.2M distribution business with three locations and 14-vehicle fleet is the canonical example.
• The rate-filing memo prompt names the actuarial credential, the ASOPs by number (56, 41, 23), and the Akur8 model output as context. Restricts SHAP/PDP fabrication; restricts Akur8 Discover from being cited as adverse-action driver; demands NAIC §4.3 and §4.4 summaries. The Q2 commercial-property base-rate revision with +4.2% statewide indication is the canonical example.
• The reason-code clause is what makes the output §4-compliant. Enumerate the variables, flag protected-class proxies, separate AI signal from human-reviewable facts. The auditable version survives the §4 exam, the bad-faith complaint, and the producer pushback.
• The role framing is the model's anchor for which discipline applies. A claims-adjuster role on a UW prompt drifts toward loss-handling; a producer role on a claims FNOL drifts toward client-relationship language; an actuary role on a renewal narrative loads ASOP language into a broker pack. Role line is structural, not stylistic.
• The prompts are templates, not formulae. The five-line preamble, the five-part body, and the reason-code clause are durable architecture; the artifact-specific context, constraints, and formats customize per workflow. L4 prompt-log retention captures every run; L3 RAG loads the retrieval sources; L2 verification catches output failures.