pre-analysis

Audits tabular data and recommends modeling choices before any fit. Produces a PreAnalysisReport JSON validating against the pack schema and a standalone HTML report.

When this skill fires

• User has tabular data + identified target, but hasn't fitted yet
• User asks for EDA, data audit, "what should I do with this data?"
• User asks whether to transform variables before modeling
• Before any fit skill — especially before linear-regression if heteroscedasticity or skew is plausible

Inputs

• --data <path> — CSV or Parquet
• --target <column> — target variable (continuous, binary, or count)
• --features <a,b,c> — comma-separated predictor columns, or all to use every non-target column
• --output <dir> — output directory

How to invoke

uv run python .agents/skills/pre-analysis/scripts/audit.py \
    --data path/to/data.csv \
    --target price \
    --features all \
    --output results/

Outputs results/pre_analysis.json (PreAnalysisReport) and results/pre_analysis.html.

Verbalising the output

Read the warning block first (flags field, severity HIGH first). Then walk through modeling_recommendations — these are the concrete actions the user should take before fitting. Do not skip the recommendations to dive into raw distributions; the recommendations are the deliverable.

If target.recommendations includes log_transform, suggest passing --log-target to the subsequent fit skill. If multicollinearity.flagged is non-empty, suggest either dropping features or using regularized-regression with ridge.

Chaining

The typical next step is one of:

• linear-regression (continuous target, OK assumptions) — pass any recommended transforms
• logistic-regression (binary target)
• regularized-regression (multicollinearity flagged)

Reference files

• references/transformations.md — when to log/sqrt/Box-Cox/Yeo-Johnson
• references/encoding.md — categorical encoding choices
• references/vif_remediation.md — what to do about multicollinearity