claude-skills
Three Claude Code skills, each a folder with a SKILL.md that Claude loads when a request matches it: Markdown to paginated PDF, EC2 offload for ML training, and a look-ahead bias audit for time-series datasets.
- Repository
- darblinko/claude-skills
- Revision
31cbad1- Measured
- 2026-10-28
- License
- Apache-2.0
Problem
Two failures that pass without an error message. Chrome's print engine clips anything that overflows the page box instead of continuing it on the next page, so a mermaid diagram taller than one page loses its lower part. In a time-series dataset, a feature built with groupby(bucket).transform("last") gives every bar the bucket's final value, so on 1-minute bars in 5-minute buckets 4 of every 5 bars read a close recorded after them.
How it works
Claude Code reads only each skill's name and description until a request matches, then loads the full SKILL.md. The PDF skill runs a script; the other two are instructions with code patterns. The renderer pipeline:
flowchart TD
md[Markdown file] --> extract[Extract mermaid fences]
extract --> html[markdown-it to HTML]
html --> chrome[Chromium at printed width]
chrome --> render[mermaid to SVG]
render --> fit{Fits the page?}
fit -- yes --> scale[Scale to fit]
fit -- too tall --> slice[Slice across pages]
fit -- far too wide --> band[Stacked bands]
scale --> pdf[PDF with header and footer]
slice --> pdf
band --> pdf
Measured results
The look-ahead skill's Python blocks, executed as written from its SKILL.md on synthetic random-walk bars. First, which bucket patterns read a later close under three bar conditions; then the correlation audit on a dataset with two planted leaks.
Run it
git clone https://github.com/darblinko/claude-skills mkdir -p ~/.claude/skills cp -r claude-skills/skills/markdown-to-pdf ~/.claude/skills/cd claude-skills uv sync && uv run pytest -qcd skills/markdown-to-pdf npm ci && npx playwright-core install chromium npm test # needs poppler-utils
Limitations
- The original daily-merge fix in the look-ahead skill shifted the date key with the values and removed no look-ahead. It is fixed, and a test keeps the original pattern as a negative case.
- The row-count bucket shift is leak-free only for bars of 1 minute or longer; on 30-second bars it reads later closes (chart above). The gap-safe variant does not.
- The look-ahead measurements use synthetic data and say nothing about effect size on a real model.
ec2-ml-trainingships no scripts; nothing in the repository runs its interface.- The PDF tests check page counts, extracted text and embedded images, not visual layout. A narrow, tall diagram is scaled up to the column width before slicing and can span many pages.