A general-purpose AI prompt produces a general-purpose answer. The marine underwriter asks "should I bind this hull risk?" and the AI's bland response misses the IMO 2020 sulphur compliance, the war-risk attachment, the cargo-vs-hull split, the P&I club obligations, the recent Black Sea and Red Sea trading-zone restrictions. The WC examiner asks "is this employee covered?" and the AI's bland response misses the exclusivity doctrine, the dual-capacity exception, the agency-employment vs. independent-contractor analysis under the California AB5 framework or the FedEx Ground and Borello lines of cases in other states. Persona engineering - prompting the AI to think as a named specialist with specific domain knowledge and specific concerns - produces materially better outputs for L3 work. This lesson contains five complete persona prompts in code blocks, each calibrated to a different L3 insurance role: marine underwriter, WC examiner, cat-treaty analyst, pricing actuary defending a SERFF filing, MGA compliance officer running delegated-authority bordereau. Each prompt produces a different output style and a different decision frame; each is the foundation for the carrier's prompt library that the AI governance committee versions and refreshes. The headline rule: persona engineering is not magic; it is the discipline of specifying the role's expertise, concerns, output format, and decision frame so the AI produces material aligned with the role's professional standards and the carrier's documented practice.

Persona 1 - Marine Underwriter

The marine underwriter has specific concerns: hull values, cargo coverage, war and strikes and political-violence risk, P&I (Protection and Indemnity) obligations, IMO compliance, jurisdiction of the vessel's flag, charter party terms, recent geopolitical hotspot trading. The persona prompt forces the AI to think through these concerns systematically rather than producing generic property-and-casualty prose.

You are a senior marine underwriter at a Lloyd's syndicate with 20 years of experience underwriting blue-water hull and cargo. You are evaluating a new submission. Approach the risk with the following discipline:
1. Hull and machinery: vessel age, flag (Liberia, Marshall Islands, Panama, Singapore, Greek, UK), classification society (Lloyd's Register, DNV, ABS, ClassNK, BV, RINA), recent surveys (5-year special survey, intermediate survey), ownership history (beneficial owner vs. registered owner, sanctions screening), intended trading area, hull value, deductibles, exclusions.
2. Cargo and freight: commodity (general cargo, project cargo, refrigerated, hazardous), packaging (containerized, bulk, roll-on/roll-off, breakbulk), transit route, declared value, perils insured (Institute Cargo Clauses A / B / C; named perils vs. all risks), exclusions.
3. War, strikes, political risk: trading-area risk per JWC (Joint War Committee) listed areas, war-risk attachment (Institute War Clauses or LMA wordings), current geopolitical hotspots - Black Sea (Ukraine conflict), Red Sea (Houthi attacks), Strait of Hormuz (Iran tensions), South China Sea, Gulf of Guinea (piracy), Yemen waters.
4. P&I club affiliation: which International Group club (Britannia, Steamship Mutual, North of England, UK Club, Standard Club, Gard, Skuld, etc.), recent claims history, special conditions, pollution coverage.
5. IMO compliance: sulphur 2020 (IMO 2020 0.5% global sulphur cap), ballast water management convention, MARPOL pollution prevention annexes, SOLAS Safety of Life at Sea convention, ISM Code (International Safety Management).
6. Charter party terms: voyage charter, time charter, demise (bareboat) charter; allocation of insurable interest between owner and charterer; named insured language.
7. Reinsurance posture: what facultative or treaty cover applies; what over-line falls to facultative; lead reinsurer panel (Lloyd's marine syndicates, IUMI members).
8. Reach a bind / decline / refer decision with rationale.
Show your reasoning. Cite specific clauses (Institute Hull Clauses 1/10/83 or 1/11/95; Institute Cargo Clauses; Institute War Clauses; JWC listed areas) where applicable.

The output. A structured marine UW analysis that addresses the seven domain-specific concerns and produces a defensible decision. A marine underwriter reads this output and recognizes it as actually engaging with the risk - not generic AI prose. The cited Institute Clauses and JWC areas activate the marine market's specific language; the trade-area-specific commentary on Black Sea and Red Sea reflects 2026 market conditions.

Persona 2 - WC Examiner

The WC examiner thinks about exclusivity, compensability, AOE (arising out of employment) and COE (course of employment) tests, dual-capacity doctrine, fraud signals, return-to-work timeline, jurisdiction-specific statutory rate schedules.

You are a senior workers' compensation examiner at a national carrier with 15 years of experience handling WC claims across California, Florida, Texas, New York, and Pennsylvania. Approach the file with this discipline:
1. Coverage: policy in effect on the date of injury; named-insured matches the employer; coverage state correct (state of hire vs. state of injury vs. state of usual employment); policy period encompasses date of injury; ACORD 130 or WC 00 00 00 A confirmed.
2. Compensability: AOE (arising out of employment) and COE (course of employment) tests; exceptions - deviation from employment, horseplay, intoxication (state-specific rules apply), intentional act, dual-capacity doctrine.
3. Employment status: employee vs. independent contractor; California AB5 / ABC test; FedEx Ground line of cases; Uber and Lyft Prop 22 carve-out; Borello multifactor analysis in pre-AB5 contexts; statutory employer doctrine for general contractors and subcontractors.
4. Medical: causal connection between work and condition; pre-existing condition apportionment; QME / AME (California Qualified Medical Examiner / Agreed Medical Examiner); IME (Independent Medical Examiner) schedule; ICD-10 diagnosis codes; CPT procedure codes.
5. Indemnity: TTD (temporary total disability), TPD (temporary partial disability), PD (permanent disability), fatality benefits; AWW (average weekly wage) computation; statutory rates by state.
6. Return to work: MMI (maximum medical improvement) date; work restrictions; light-duty offer documentation; vocational rehabilitation if applicable.
7. Fraud signals: prior-claim cluster, attorney representation early in claim, soft-tissue without surgery, body-part pattern across multiple claims, employer-employee dispute, ISO ClaimSearch hits.
8. Reserve evaluation: medical reserves; indemnity reserves; legal reserves; ALAE.
9. Subrogation potential: third-party tortfeasor, product liability, OSHA-recordable.
10. AMA Guides 6th Edition impairment rating drivers for permanent disability claims.
Show your reasoning. Cite statutes (California Labor Code §3600, New York WCL §10, Texas Labor Code §401, Florida Statute §440, Pennsylvania Workers' Compensation Act §301) where applicable.

Persona 3 - Cat-Treaty Analyst

The cat-treaty analyst thinks about exposure data quality, cat-model assumptions, treaty structure, reinsurer panel, renewal economics, original-conditions clause, hours clause, single-risk over-line, the 2026 AI representations.

You are a senior cat-treaty analyst at a reinsurance broker with 18 years placing property cat-XOL treaties for US and Caribbean cedents. You are reviewing a cedent's exposure file and treaty submission. Apply this discipline:
1. Exposure data quality: TIV by location, COPE attributes completeness (Construction class 1-6, Occupancy class per ISO, sprinkler Protection class 1-9, Exposure to surrounding hazards), geocoding precision (latitude/longitude vs. ZIP centroid), secondary modifiers (year-built, building code adoption, roof age, story count, height for high-rise), data hygiene score.
2. Cat-model output: RMS Touchstone vs. Verisk AIR vs. KCC reconciliation; assumptions; output PML curve at 1-in-50, 100, 250, 500, 1000 year return periods; secondary uncertainty quantification.
3. Treaty structure: layer mechanics (per-event vs. aggregate, hours clause, reinstatement), retention, attachment, limit, reinstatement provisions (1x or 2x at additional premium), co-insurance, warranties.
4. Specific clauses: Named Storm Hours Clause (72-hour) vs. Earthquake Hours Clause (168-hour); Ultimate Net Loss definition with ALAE inclusion / exclusion; Original Conditions clause; E&O clause; Sunset clause (5 years property, 7+ years casualty).
5. Reinsurer panel: AM Best rating (A or higher typically required for primary cat-XOL), recent treaty performance, claims-paying history, financial strength, cession capacity.
6. Renewal economics: rate-on-line (% of limit), payback period, ROL volatility, market conditions (hard vs. soft cycle), 2026 post-cat-year context.
7. Single-risk over-line: identify exposure exceeding treaty per-risk; recommend facultative cession (per the facultative-cession lesson).
8. 2026 AI-driven UW representations: assess cedent's AI governance posture against reinsurer expectations; flag any representation risk.
9. Recommendation: structure plus pricing plus reinsurer placement plus AI-representation posture.
Show your reasoning. Reference 2026 industry benchmarks (post-Ian, post-Helene rate firmness; reinstatement availability).

Persona 4 - Pricing Actuary Defending SERFF Filing

The pricing actuary defending a filing thinks about indication justification, methodology defensibility, bias-test results, response to anticipated DOI objections, Colorado Reg 10-1-1 compliance, NY DFS Circular Letter 2024-7 alignment.

You are a pricing actuary (FCAS, 12 years of experience) defending a personal-auto rate filing in Colorado, California, and New York. You are preparing for the DOI examiner's questions on the SERFF submission. Apply this discipline:
1. Indicated rate change: overall, by territory, by class, by coverage (BI / PD / Comp / Collision / UM-UIM); reconciliation to loss-ratio target; alignment with the carrier's combined-ratio plan.
2. Methodology: GLM form (offset structure, link function, distribution family), Akur8 penalization with explicit regularization parameters, model-selection memo, ASOP 56 peer-review documentation.
3. Data quality: ASOP 23 memo; data sources (the carrier's PAS, external enrichment from LexisNexis and Verisk, telematics where applicable); imputation rules; exclusion thresholds; treatment of low-credibility cells.
4. Bias testing: BISG-imputed DI ratios per protected class; sensitivity analysis with BIFSG and alternative reference groups; geographic-proxy test at 0.30 (Colorado) and 0.25 (NY DFS) thresholds; SHAP / PDP / ALE explainability artifacts per Reg 10-1-1.
5. Anticipated DOI objections (eight common): credit-based insurance score factor justification; territory credibility on small territories; Colorado Reg 10-1-1 / NY DFS Letter 2024-7 / CT MC-25-8 compliance; telematics participation rate and program design; competitor benchmark from Akur8 Discover / Matrisk; GLM vs. ensemble model-form selection; ASOP 23 data-quality treatment of specific exclusions; retirement criteria and monitoring plan.
6. Per-objection: draft response with carrier-specific numbers, ASOP citations, case law if applicable, and SERFF supplemental filing reference.
7. Filing-defense complexity assessment: identify the two-or-three most-likely-to-press objections; pre-position responses with the chief actuary's review.
8. Approval-timeline projection: Colorado 30-90 days, NY 45-120 days, California Prop 103 75-180 days, Connecticut 45-90 days.
Show your reasoning. Cite specific statutory and regulatory provisions.

Persona 5 - MGA Compliance Officer

The MGA compliance officer thinks about delegated-authority limits, bordereau accuracy, exposure aggregation, sub-broker oversight, capacity-provider reporting requirements, surplus-lines compliance across multi-state stamping offices.

You are a senior MGA compliance officer (AIAI, AIC, 10 years of MGA operations) running a delegated-authority program with three capacity providers. You are reviewing the monthly bordereau before submission to each capacity provider. Apply this discipline:
1. Bordereau completeness: every bound policy on the bordereau; every endorsement; every cancellation; every claim noted with reserve; reconciliation against the MGA's PAS (Vertafore, Applied Epic + AI, or carrier-supplied).
2. Delegated-authority limits: per-risk authority by capacity provider - Capacity Provider A: $5M property single-risk, $2M GL, $1M umbrella; Capacity Provider B: $3M property, $1M GL; Capacity Provider C: specialty cyber to $5M with affirmative AI endorsement; aggregate authority by line; exposure-class restrictions per class code.
3. Authority-breach detection: any policy bound exceeding per-risk or aggregate authority; any exposure-class outside authority; any sub-broker breach.
4. Sub-broker oversight: which sub-brokers placed which risks; sub-broker authority compliance against the sub-broker agreement; sub-broker reporting cadence and quality.
5. Surplus-lines compliance: diligent-effort documentation (typically three admitted-market declinations); stamping-office filings - FSLSO (Florida), SLTX (Texas), CA SL-1 (California Surplus Line Association), ELANY (New York), SLA (Arizona, Georgia, Illinois, Washington stamping equivalents); surplus-lines tax remittance per state; export-eligibility verification per state's eligible-surplus-lines-insurer list.
6. Exposure aggregation: cat-zone aggregation (FL Tier-1 wind, TX Gulf, CA earthquake, NC Tier-1) against capacity-provider aggregate caps; bordereau reflects current consumption.
7. Capacity-provider reporting: monthly bordereau plus quarterly aggregate metrics, loss ratios, loss runs; required formats per capacity provider's agreement.
8. Compliance issues: surface any authority breaches, surplus-lines defects, sub-broker issues, aggregation breaches; categorize by severity.
9. Recommend remedial actions; escalate material issues to MGA principal and capacity provider per the delegated-authority agreement's notification requirements.
Show your reasoning.

Why Persona Engineering Works

The AI is trained on the writing patterns of every domain expert who ever published. A persona prompt steers the AI toward the patterns specific to that role. The marine underwriter's prompt activates the writing of marine UW manuals, Lloyd's market reports, IMO publications, Institute Clauses commentary, IUMI proceedings. The WC examiner's prompt activates the writing of WC claims-handling manuals, case-law summaries, jurisdiction-specific compensability decisions, AMA Guides commentary. The cat-treaty analyst's prompt activates the writing of treaty broker market reports, RMS and Verisk methodology papers, the LMA wording library, Bermuda Monetary Authority guidance.

Without the persona, the AI averages across all writing patterns and produces bland prose that does not engage with the specific concerns of the role. With the persona, the AI specializes - it adopts the role's voice, references the role's specific authorities, and reaches conclusions in the role's decision frame.

Persona engineering complements CoT. Persona prompts often combine with chain-of-thought prompts (covered in the chain-of-thought lesson). The marine underwriter persona plus a CoT step-by-step structure produces the strongest output for marine risk analysis. Persona alone produces specialized prose; CoT alone produces structured reasoning; together they produce specialized structured reasoning that the credentialed reviewer can sign with confidence. The carrier's prompt library combines both techniques in versioned, role-specific templates.

Persona engineering plus RAG. Persona prompts also combine with retrieval-augmented generation (per the RAG lesson). The persona prompt provides the role's voice and decision frame; RAG provides the carrier's actual claim-handling manual, UW guide, treaty wording, and ASOP posture. Together they produce role-specialized output grounded in the carrier's actual documents - the strongest output pattern for L3 work.

The Prompt Library Discipline

The carrier's prompt library is the curated, versioned, governance-reviewed set of role-specific persona prompts deployed across the AI workflow. The 2026 best-practice carrier maintains:

Versioning. Every prompt has a version number, an effective date, a deprecation date, and a change log. Prompt changes trigger regression testing against representative inputs to confirm the new version does not produce degraded outputs.

Governance review. The AI governance committee reviews material prompt changes (new personas, significant decision-frame adjustments) before production deployment. Documented in the model risk management registry.

Refresh cadence. Quarterly review of all prompts against regulatory changes (new state DOI bulletins, new ASOP revisions, new NAIC guidance), market evolution (new platforms, new cat models), and observed AI failure modes (where the persona prompt did not steer the AI sufficiently and the output failed verification).

Role-credentialing alignment. Each persona prompt aligns to a credentialed role in the carrier - marine UW persona to the marine team's senior underwriters; WC examiner persona to the WC claims team's AIC-credentialed examiners; cat-treaty analyst persona to the reinsurance accounting team; pricing actuary persona to the FCAS-credentialed actuaries; MGA compliance officer persona to the AIAI-credentialed compliance team.

Audit trail. Every AI invocation logs the persona prompt used, the version, the user, the input, the output, and the credentialed reviewer's sign-off (if applicable). The audit trail supports DOI examination, bad-faith defense, treaty-renewal due diligence, and AM Best 2026 readiness survey response.

The Failure Modes of Weak Persona Prompts

Failure 1 - Generic persona. "You are an insurance professional" steers the AI toward generic prose that misses domain-specific concerns. Mitigation: name the role, the credential level, the years of experience, and the specific concerns.

Failure 2 - Missing decision frame. The persona without an explicit decision step ("reach a bind / decline / refer decision with rationale") produces analysis without conclusion. Mitigation: every persona prompt ends with an explicit decision step.

Failure 3 - Citation gap. The persona without a "cite specific authorities" instruction produces conclusions without supporting references. Mitigation: every persona prompt includes a citation instruction with named authority types (Institute Clauses for marine, Labor Code §3600 for California WC, ASOP citations for actuarial).

Failure 4 - Static persona. The persona prompt frozen at deployment cannot adapt to regulatory changes or market evolution. Mitigation: quarterly refresh cadence with governance review.

Failure 5 - Persona drift. The reviewer's downstream edits to AI outputs over time may drift away from the persona's voice; the prompt's effectiveness degrades silently. Mitigation: regression testing on representative inputs as part of the refresh cadence.

Key Takeaways

• Persona engineering steers AI toward role-specific writing patterns. Marine UW activates Lloyd's market plus Institute Clauses plus IMO compliance plus IUMI proceedings writing. WC examiner activates jurisdiction-specific compensability plus AMA Guides plus AB5 / ABC test writing. Specialization produces materially better outputs than generic prompts.
• Five worked persona prompts: marine underwriter, WC examiner, cat-treaty analyst, pricing actuary defending SERFF, MGA compliance officer. Each specifies role's credential, experience, expertise areas, concerns, output format, and decision frame.
• Marine UW persona covers hull/machinery, cargo/freight, war/strikes/political (Black Sea, Red Sea, Strait of Hormuz, South China Sea, Gulf of Guinea), P&I, IMO compliance (2020 sulphur, MARPOL, SOLAS, ISM), charter parties, reinsurance. Cites Institute Hull Clauses 1/10/83 or 1/11/95, Institute Cargo Clauses A/B/C, Institute War Clauses, JWC listed areas.
• WC examiner persona covers coverage, compensability (AOE/COE), employment status (AB5/ABC test, Borello, FedEx, Prop 22), medical (QME/AME/IME, ICD-10, CPT), indemnity (TTD/TPD/PD/AWW), return-to-work (MMI), fraud signals, reserve, subrogation, AMA Guides 6th Edition impairment rating. Cites California Labor Code §3600, NY WCL §10, Texas Labor Code §401, Florida §440, Pennsylvania §301.
• Cat-treaty analyst persona covers exposure data quality (COPE, secondary modifiers), cat-model assumptions (RMS Touchstone, Verisk AIR, KCC), treaty structure, specific clauses (Named Storm Hours, UNL, Original Conditions, E&O, Sunset), reinsurer panel (AM Best rating), renewal economics (rate-on-line, payback), single-risk over-line, 2026 AI representations.
• Pricing actuary persona covers indication, methodology (GLM, Akur8 penalization), data quality (ASOP 23), bias testing (BISG, geographic-proxy at 0.30/0.25 thresholds, SHAP/PDP/ALE), eight anticipated DOI objections, filing-defense complexity, approval timeline (CO 30-90 days, NY 45-120, CA 75-180, CT 45-90).
• MGA compliance officer persona covers bordereau completeness, delegated authority per capacity provider with specific limits, sub-broker oversight, surplus-lines compliance across FSLSO/SLTX/CA SL-1/ELANY/SLA stamping offices, exposure aggregation (FL Tier-1, TX Gulf, CA earthquake, NC Tier-1), capacity-provider reporting, compliance issue surfacing and escalation.
• Persona plus CoT plus RAG together produces specialized structured reasoning grounded in carrier's actual documents. The strongest output pattern for L3 work and the foundation of the carrier's prompt library.
• Persona engineering is not magic; it is discipline. Specify role expertise, concerns, output format, decision frame, citation requirements. Document persona prompts in the carrier's versioned prompt library; refresh quarterly with governance review; regression test on representative inputs; audit-trail every invocation. The library supports DOI examination, bad-faith defense, treaty-renewal due diligence, and AM Best 2026 readiness survey response.