Data

Data Skill

Two modes. Match the request to a section.

---

A. Data Analysis

Every analysis starts with the decision it supports, works backward to evidence required, then touches the data. Analysis without a decision is arithmetic.

Phase 1: FRAME

Establish what decision this analysis supports.

1. Identify the decision, decision-maker, options, and default action if no analysis is done.
2. If the user cannot articulate a decision, ask: "What will you do differently based on this analysis?" If exploratory, switch to Exploratory Mode (apply rigor gates, make no causal claims).
3. Define evidence requirements: what evidence favors each option, minimum threshold for changing the default, deal-breakers.
4. Save analysis-frame.md.

Gate: Decision identified, options enumerated, evidence requirements saved.

Phase 2: DEFINE

Lock metric definitions before loading data. Defining after seeing data enables cherry-picking.

For each metric: name, exact formula (numerator/denominator), population (included/excluded), time window, segments. For comparisons: define groups and verify fairness.

Save metric-definitions.md. Definitions are locked once Phase 3 starts. If data reveals a definition is unworkable, return here, update, and document the change.

Gate: All metrics defined with formulas and populations.

Phase 3: EXTRACT

Load data. Assess quality. No interpretation.

1. Detect tools: try import pandas; fall back to csv.DictReader + statistics.
2. Profile: row count, column types, missing values, date range, distribution stats.
3. Quality checks (load references/rigor-gates.md Gate 1):

4. Save data-quality-report.md.

Gate: Data loaded, quality assessed, failures documented as limitations.

Phase 4: ANALYZE

Compute metrics per Phase 2 definitions. Report confidence intervals, not point estimates.

1. Compute using exact formulas. Wilson score CI for proportions.
2. Fairness gate (comparisons): same time window, same population, confounders documented, survivorship checked (load references/rigor-gates.md Gate 2).
3. Multiple testing (6+ comparisons): apply Bonferroni (threshold = 0.05/N). Report all segments tested (Gate 3).
4. Practical significance: report effect size alongside statistical significance. Base-rate context ("from 2.1% to 2.3%", not "+10% lift") (Gate 4).
5. Save analysis-results.md.

Gate: All metrics computed. Rigor gates applied.

Phase 5: CONCLUDE

Lead with insights. Return to the decision.

1. Headline finding: one sentence addressing the Phase 1 decision.
2. Supporting evidence: primary metric with CI, secondary metrics, segment breakdowns.
3. Limitations: wide CIs are the finding, not a formatting problem.
4. Decision mapping: does evidence meet threshold? Deal-breakers triggered? Recommended action? Additional data needed?
5. Save analysis-report.md (load references/output-templates.md for analysis-type templates).

Gate: Report saved with headline, limitations, recommendation tied to decision.

Error Handling (Data Analysis)

---

B. Reference Enrichment

Enrich agent/skill reference files from Level 0-2 to Level 3+, or decompose bloated body files by extracting domain content into references.

Phase 0: DECOMPOSE (when --decompose or "extract references")

Extract domain-heavy content from a bloated SKILL.md into reference files.

1. Run python3 scripts/detect-decomposition-targets.py --skill {name} (or --agent).
2. If no extractable blocks, report "nothing to decompose" and stop.
3. Snapshot: cp {path} /tmp/decomp-before-{name}.md.
4. For each block: create reference file, remove from body (MOVE, not copy), add loading table entry.
5. Retain in body: frontmatter, overview, phase workflow, loading table, error handling.
6. Validate: python3 scripts/validate-decomposition.py --before /tmp/decomp-before-{name}.md --after {path} --refs {refs_dir}/.
7. If fails: restore from snapshot. If passes: python3 scripts/validate-references.py --skill {name}.

Load references/decomposition-prompt.md for the autonomous decomposition prompts.

Gate: Validation passes. Body reduced. All extracted content in references.