Diagnostics

Skill WindcleaverDev/regkit/.agents/skills/diagnostics

Statistically rigorous regression skills for Claude — Python computes, the model narrates.

Install

npx -y skills add WindcleaverDev/regkit --skill diagnostics

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Run the full diagnostic battery on a fitted regression model — assumption tests (linearity, homoscedasticity, normality of residuals, independence, multicollinearity), influence analysis (leverage, Cook's distance, DFFITS), and bias/variance assessment via cross-validation. Produces a structured DiagnosticsReport with status flags and ranked remediation recommendations, plus a standalone HTML report. Use this skill whenever a regression has been fitted and the user wants to know if they can trust it, asks "are the assumptions met?", "is this model overfit?", "is this overfitting or underfitting?", or wants to identify influential outliers. Pair with the linear-regression or regularized-regression skills.

SKILL.md

2.4 KB, 399 tokens by cl100k_base, as published. Nobody here has run it

diagnostics

Runs the full diagnostic battery on a fitted regression model. Input is the JSON output of a fit skill (e.g. LinearRegressionReport) plus the original data.

When this skill fires

User has fitted a regression (via linear-regression or similar) and asks about model quality
User asks "is this overfitting?", "are the assumptions met?", "is there multicollinearity?"
User asks to identify influential observations, outliers, or high-leverage points
The fit skill recommends running diagnostics

Inputs

--fit-report <path> — JSON of a LinearRegressionReport (or compatible)
--data <path> — the original data file (must contain the same target & features used in the fit)
--output <dir> — output directory

Optional:

--cv-folds <int> — folds for cross-validation (default 5)
--learning-curve — include learning curve data in the report (slower)
--test-split <float> — held-out fraction for train/test gap (default 0.2)

How to invoke

uv run python diagnostics/scripts/diagnose.py \
    --fit-report results/report.json \
    --data data/houses.csv \
    --output results/

Outputs results/diagnostics.json (DiagnosticsReport) and results/diagnostics.html.

Verbalising the output

Read the verdict.headline first. Then surface verdict.top_issues in order — each maps to one or more flags. For each flag, the report's recommendations field has the actionable next step. Do not list every assumption check unless asked; lead with what's broken or marginal.

Reference files

references/plots_guide.md — what each diagnostic plot should look like
references/remediation.md — what to do when each assumption fails