Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Platform-neutral analytical skill that profiles messy data, selects case-adaptive methods, and produces source-backed visual reports for high-stakes decisions
triggers
["analyze this dataset and build an evidence-based report","run a high-stakes decision analysis with data quality gates","create an evidence intelligence report from this data","perform adaptive analytics with diagnostic predictive or prescriptive routing","validate this data and choose the right analytical method","generate a decision intelligence brief with uncertainty bounds","profile data quality and route to appropriate analysis method","build a reproducible evidence report with data lineage"]
A platform-neutral, evidence-constrained analytical system that transforms ambiguous questions into reproducible evidence products. It profiles data quality, selects case-adaptive analytical methods (descriptive, diagnostic, predictive, prescriptive), and produces source-backed reports with explicit uncertainty and claim boundaries.
What It Does
Instead of forcing every dataset through fixed pipelines, this system:
Gates data quality before analysis (detects missing, duplicates, leakage, grain mismatches)
Routes adaptively to descriptive, diagnostic, predictive, or prescriptive methods based on question + data
project:name:"customer-churn-analysis"question:"Which customers are at risk of churning in next 90 days?"decision_owner:"Head of Retention"evidence_contract:source:"data/customer_events.csv"grain:"customer_id"time_field:"event_date"target_field:"churned"horizon_days:90data_quality:missing_threshold:0.15duplicate_check:trueleakage_detection:trueprivacy_scan:truerouting:force_descriptive:trueenable_diagnostic:trueenable_predictive:trueenable_prescriptive:falseoutputs:evidence_report:"outputs/evidence_report.md"decision_brief:"outputs/decision_brief.md"figures_dir:"outputs/figures/"reproducibility_package:"outputs/reproducibility.zip"
# Use when data quality blocks analysis
gate = DataQualityGate(df, contract)
report = gate.profile()
if report.status == "blocked":
evidence_request = {
"status": "evidence_request",
"reason": report.blocking_reason,
"required_corrections": report.required_corrections,
"resubmit_with": report.corrected_contract
}
# Stop here, do not proceed to analysisreturn evidence_request
Pattern 2: Negative Validation → Do Not Deploy
# Use when predictive model fails validation
predictor.fit(df_train)
validation = predictor.validate(df_test)
if validation.deployment_status == "do_not_deploy":
decision_brief = {
"status": "negative_validation",
"evidence": validation.evidence_report_link,
"blocking_issue": validation.blocking_reason,
"alternatives": ["collect_more_data", "revise_estimand", "stop"]
}
# Do not deploy, document negative resultreturn decision_brief
Pattern 3: Evidence Sufficient → No Decision Layer Needed
# Use when question is purely evidentialif question_type == "evidence_request":
# Generate evidence report only
evidence = generate_evidence_report(results)
# Do NOT force a decision briefreturn {"evidence_report": evidence, "decision_brief": None}
Pattern 4: Adaptive Route Composition
# Use when multiple routes are justifiedif data_characteristics.supports_multiple_routes():
route = {
"primary": "descriptive", # Always first"additional": ["diagnostic", "predictive"], # Add if justified"excluded": ["prescriptive"], # Not enough for action"reason": "Insufficient alternatives and constraint data"
}
Troubleshooting
Data Quality Gate Blocks Analysis
Problem: status: "blocked" with reason: "grain_violation"
Solution: Ensure your data contract matches actual data structure
# Check grain uniquenessprint(f"Unique grain values: {df[grain_field].nunique()}")
print(f"Total rows: {len(df)}")
# If not unique, identify duplicates
dupes = df[df.duplicated(subset=[grain_field], keep=False)]
print(dupes)
# Fix contract or deduplicate explicitly
Missing Field Errors
Problem: KeyError: 'target_field'
Solution: Verify all contract fields exist
contract_fields = [contract["grain"], contract["time_field"], contract["target_field"]]
missing = [f for f in contract_fields if f notin df.columns]
if missing:
print(f"Missing fields: {missing}")
print(f"Available columns: {df.columns.tolist()}")
Route Selection Returns "descriptive_only"
Problem: Expected predictive route but got descriptive only
Solution: Check data volume and target prevalence
print(f"Rows: {len(df)}")
print(f"Target prevalence: {df[target].mean():.3f}")
print(f"Positive cases: {df[target].sum()}")
# Predictive requires minimum sample size and events# Typically: n > 500 AND positive_cases > 50
from sklearn.calibration import CalibratedClassifierCV
# Recalibrate model
calibrated = CalibratedClassifierCV(model, method='isotonic', cv=5)
calibrated.fit(X_train, y_train)
# Or document as limitation
limitation = {
"issue": "poor_calibration",
"metric": f"slope={calibration_slope:.2f}",
"boundary": "Use for ranking only, not absolute probabilities"
}
Decision Brief Generation Fails
Problem: decision_status: "no_decision_ready"
Solution: This is often correct — not every analysis should produce a decision
# Check if decision layer is actually justifiedifnot (
feasible_alternatives_exist and
constraints_defined and
decision_owner_identified and
reversal_conditions_specifiable
):
# Correctly stop at evidence layerprint("Evidence report is terminal product")
