| name | execute-consulting-analysis |
| description | Convert approved source evidence into transparent derived conclusions through calculations, scoring, sizing, benchmarking, experiments, diagnostics, business cases, models, scenarios, or sensitivity analysis. Use when a consulting hypothesis depends on a material transformation rather than a reported source fact. Return auditable inputs, method, assumptions, checks, result, sensitivity, and limits to the central lead; do not adjudicate the overall answer or write the storyline. |
Execute Consulting Analysis
Create the transparent transformation between approved evidence and a derived result. Preserve the central lead's bounded test; do not decide the overall answer.
Use this skill only when the result materially depends on a calculation, sizing, score, benchmark transformation, experiment, diagnostic, business case, model, scenario, multi-source reconciliation, or sensitivity—not to repeat a reported number.
Run the analysis loop
- Restate the test. Preserve the hypothesis, decision use, method boundary, and support/refute/inconclusive conditions. State what the analysis can and cannot establish.
- Lock material inputs. Record decisive source/extract, version, fields, population, period, units, denominators, and transformations. Reuse the evidence skill's canonical
minimum_access_request when a decision-changing input is inaccessible; do not reproduce a competing field specification.
- Choose a readable method. Explain the formula, model, comparison, score, experiment, or solver; expose assumptions and decision-changing rules.
- Run relevant checks. Reconcile totals, units, denominators, missingness, duplicates, joins, ranges, signs, formulas, and benchmark comparability. Test calibration, leakage, bias, confounding, or model validity when applicable.
- Test sensitivity and rivals. Vary inputs/assumptions capable of changing the decision and test whether a simpler rival fits.
- Rerun only affected logic when a qualified input changes; show the exact before/after input and result delta.
- Return, do not adjudicate. State result, method, checks, sensitivity, limits, suggested hypothesis status, and conclusion boundary.
Apply decision rules
- A weighted score is not self-validating. Check construct fit, weighting rationale, missing-data policy, comparable scales, threshold logic, and weight sensitivity.
- A benchmark needs aligned population, scope, period, geography, currency, denominator, maturity, and method. Show a common-basis bridge or transparent bounded normalization when exact conversion is impossible;
not comparable alone is not a result.
- For utilization, capacity, throughput, or constraint claims, compare the mechanism-matched views that may bind: nominal availability, staffed/startable window, cycle/start-slot capacity, and complementary-resource load. Reconcile observed throughput with the maximum implied starts/cycles; quantify any inconsistency.
- A model failing a required check cannot support its conclusion. A reversing sensitivity remains visible.
- A proxy is labeled and states which direct claim it cannot replace.
- For each derived benefit, cost, or exact decision threshold, label provenance as
evidenced or judgment and state additive versus overlapping treatment. Do not turn a category share into causal addressability by multiplying it by a total unless the mechanism is evidenced; if used as a scenario, state the assumption and range. Untested option efficacy remains a hypothesis.
Return one compact note
Include test/decision use; input locators and definitions; method/formula; assumptions; checks and failures; result with units/scope; sensitivity and rival; limitation/blocked item; suggested support, refute, or inconclusive; any inherited restrictions; and, for reruns, the prior-result locator and before/after delta. Mark preliminary, directional, or non-comparable results honestly.
Load detail only when needed
- Read analysis-workflow.md for multi-step, rerun, or consequential models—not for a short transparent calculation.
- Read method-gates.md only for a specialized method needing an explicit feasibility/validity gate.
- Read quality-bar.md for consequential handoff or failed review.
- Reusable or consequential models may need organization-specific frozen-input, permission, lineage, and handover controls; do not manufacture that overhead for ordinary analysis.
