Explore AI Agent Skills & Claude Prompts

Phase 6 Commit ι requirement. When a model produces an allocation or policy recommendation, the report must include DALY-averted figures alongside cases-averted unless the modeler explicitly justifies their absence. DALYs (disability-adjusted life-years) are the standard metric for Global Fund / GBD / WHO health-allocation analyses; cases-averted alone treats a 6-month-old's averted infection identically to a 30-year-old's, and dramatically under-weights interventions that target high-mortality subpopulations (SMC for U5 children, vaccines for newborns, ART for advanced HIV). Trigger phrases include "DALY", "disability-adjusted life-years", "cost per DALY", "DALY-averted", "GBD weight".

uncertainty-quantification

Contract for STAGE 5b UNCERTAINTY. Modeler exposes an outcome_fn in models/outcome_fn.py that takes sampled parameter values (from the effect-size priors registry) and returns decision outputs as a dict. The UQ script samples N draws from each registered parameter's prior, runs outcome_fn once per draw, and aggregates per-output 95% credible intervals into uncertainty_report.yaml. Scalar outputs (DALYs, costs, burden) get CIs; categorical outputs (per-archetype package choice, per-LGA assignment) get stability distributions. The writer must report these posterior-derived CIs as the primary uncertainty claim — NOT an ensemble ±X% perturbation. The gate blocks ACCEPT without uncertainty_report.yaml. Use when writing outcome_fn, interpreting posterior CIs, or deciding whether to use cloud compute (see cloud-compute skill). Trigger phrases include "uncertainty quantification", "UQ", "propagate priors", "outcome_fn", "posterior CI", "credible interval", "ensemble uncertainty".

allocation-cross-validation

Phase 6 Commit κ requirement. When a model produces an allocation, the allocation rule itself must be cross-validated under spatial holdout — not just the underlying calibration. A k-fold leave-one-archetype-out test answers "would my optimizer's recommended package for archetype A change if I had calibrated on the other 21 archetypes only?" Without this test, the optimization may have over-fit to specific in-sample EIRs. Trigger phrases include "allocation cross-validation", "leave-one-archetype-out", "allocation generalizes", "allocation robustness", "spatial holdout".

basic-epi-modeling

This skill should be used when the user asks about "disease burden", "DALYs", "population at risk", "transmission routes", "epidemic vs endemic", "compartmental model basics", "SI SIS SIR model selection", "incidence vs prevalence", "WAIFW matrix", or needs foundational guidance on setting up a new infectious disease model from scratch.

malaria-modeling

Disease-specific modeling guide for P. falciparum malaria. Covers transmission dynamics, intervention mechanisms, known modeling pitfalls, calibration targets, cost-effectiveness benchmarks, and published benchmark models. Use when the research question involves malaria incidence, prevalence, transmission, interventions, or resource allocation. Trigger phrases include "malaria", "PfPR", "ITN", "bednet", "IRS", "indoor residual spraying", "SMC", "seasonal malaria chemoprevention", "MDA", "mass drug administration", "Global Fund malaria", "malaria resource allocation", "malaria optimization".

mechanistic-vs-hybrid-architecture

Phase 7 Commit ν guidance. Teaches the modeler when to use full-mechanistic (EMOD/malariasimulation/Starsim with full immunity dynamics), hybrid (mechanistic calibration + published-RCT intervention multipliers), or analytical (regression overlays) for budget-allocation analyses. The 1935 malaria run wasted 3 rounds reaching the hybrid framing because no skill taught this decision upfront. For allocation analyses where intervention effects are dominated by published meta-analyses, START WITH HYBRID. Trigger phrases include "mechanistic vs hybrid", "ABM vs analytical", "intervention effect calibration", "domain-implausible ICER", "LLIN ICER too high".

model-fitness

Framework for evaluating whether a model is fit for its stated purpose and intended audience. Covers audience-specific requirements, structural gap detection, the "simpler question" test, and Level escalation criteria. Use when deciding whether to accept a model, escalate complexity, or declare scope. Complements modeling-strategy (which covers model selection) by focusing on post-build evaluation. Trigger phrases include "fit for purpose", "is this model good enough", "should we accept", "audience requirements", "structural gap", "escalate complexity", "Level 2", "does this answer the question".

parameter-estimation

This skill should be used when the user asks about "fitting a disease model", "parameter estimation", "least squares fitting", "maximum likelihood estimation", "MLE", "Bayesian inference for epidemiology", "estimating R0 from data", "chain binomial model", "TSIR model", "time series SIR", "model calibration", "stochastic epidemic model", "measurement error", "demographic stochasticity", or needs to connect a compartmental model to observed data.

sensitivity-analysis-remediation

Phase 10 ω — interpretation and remediation of `models/sensitivity_analysis.yaml` verdicts (ROBUST / SENSITIVE / UNSTABLE). Phase 8 π made the artifact required and Phase 9 τ pushed it earlier in the pipeline. This skill teaches the modeler what each verdict means operationally and which escalation paths are available when verdict=UNSTABLE — the failure mode of the 0013 malaria run (RIG-003). Trigger phrases include "sensitivity UNSTABLE", "sensitivity SENSITIVE", "RIG-003", "perturbation flips", "primary_recommendation_changes", "narrow CI", "scope-declare sensitivity".

sir-elaborations

This skill should be used when the user asks about "SEIR model", "MSIR model", "SIRS model", "adding compartments", "age structure in models", "WAIFW matrix", "contact heterogeneity", "20/80 rule", "contact matrices", "calculating R0", "host-vector model", "multi-strain dynamics", "cross-immunity", or needs to extend a basic SIR model with additional complexity.

surveillance

This skill should be used when the user asks about "disease surveillance", "underreporting", "burden of illness pyramid", "reporting lags", "nowcasting", "epidemic forecasting", "forecast evaluation", "genomic epidemiology", "phylogenetic analysis", "molecular clock", "Google Flu Trends", "syndromic surveillance", "sentinel surveillance", or needs to work with surveillance data, forecast disease trajectories, or integrate genomic data with transmission models.

vaccination

This skill should be used when the user asks about "vaccination model", "herd immunity threshold", "critical vaccination coverage", "vaccine efficacy in models", "leaky vaccine", "all-or-nothing vaccine", "waning immunity", "pulse vaccination", "ring vaccination", "honeymoon effect", "age-shift effect", "disease eradication", "elimination vs eradication", or needs to incorporate vaccination into a compartmental disease model.