Using This Repository with an LLM

This repository provides context files, references, and templates for LLM-assisted work across multiple tools.

Which Tool, Which Approach

Tool Setup
Claude Code (CLI) Uses CLAUDE.md files automatically. See claude-code-setup.md.
Copilot / ChatGPT (chat) Paste the preamble from copilot-primer.md, then paste your code or txtarchive.
Any LLM (general) Attach or paste the relevant files from this repo as context. Follow the diff-first workflow below.

What’s Inside (Intended for LLM Context)

“Project Instructions” to Paste at the Start of an LLM Session

Copy this block into the system or first user message.

You are assisting with data analysis and document authoring.
Rules:
- Use ONLY the attached files plus official documentation if needed.
- For R code: use tidymodels (not caret), pacman::p_load(), event_level = "second".
- For Quarto: maintain valid YAML and Pandoc syntax. Assume Quarto >= 1.7.
- Preserve cross-references (@fig-*, @tbl-*, @sec-*) and citation keys (@citekey).
- Never invent references -- use only citations from the provided .bib file.
- Prefer minimal diffs. Show a unified diff for each change.
Output format for edits:
1) Short rationale (1-3 bullets)
2) Unified diff
3) Full updated file content ONLY if I request it

Prompt Cookbook

General Editing (Quarto)

  • “In _quarto.yml, enable embed-resources for HTML and add a dark theme; keep PDF unchanged.”
  • “Add a callout ‘Limitations’ after the Discussion section; show a unified diff.”

Data Analysis (R / tidymodels)

  • “Review the recipe in 310-Baseline.ipynb. Are there any preprocessing steps missing for a random forest?”
  • “Add SHAP variable importance plots to 320-Tuned-RF.ipynb using kernelshap + shapviz.”
  • “Create a Table 1 using gtsummary::tbl_summary(), stratified by death_flag.”

Research Paper

  • “Use csl/new-england-journal-of-medicine.csl and convert raw references into proper citekeys using refs/library.bib.”
  • “Create a templates/journal-reference.docx style map and adjust headings per IMRAD structure.”

Interactivity

  • “Turn analysis.qmd into a browser-only interactive doc with Plotly; make it self-contained.”
  • “Convert report.qmd to Shiny: add server: shiny, a slider input, and rendered plot.”

Citations & Zotero

  • “Wire Zotero via Better BibTeX: assume refs/library.bib is kept updated. Replace author-year mentions with @citekey syntax.”

Edit Workflow (Diff-First)

  1. You: request a change (file + intent).
  2. LLM: returns a unified diff (plus a brief rationale).
  3. You: approve or refine.
  4. LLM: provides the updated file content (or you apply the diff).
  5. You: render locally and commit.

This keeps changes auditable and reduces the chance of the LLM silently breaking something.

Conventions

  • Quarto version >= 1.7.
  • Citations: CSL + .bib, with link-citations: true.
  • Labels: fig-, tbl-, eq-, sec- prefixes; reference with @label.
  • R code: tidymodels ecosystem, event_level = "second", pacman::p_load().
  • Analysis files: NNN-Description naming (see copilot-primer.md for the numbering scheme).
  • Accessibility: add fig-alt text and meaningful tbl-cap captions.

Repository Layout

llmcheatsheets/
├── guides/                  # This file, Claude Code setup, Copilot primer
├── skills/                  # Quarto skill reference
├── cheatsheets/             # Rendered Quarto cheatsheet (.qmd + .html)
├── templates/               # Starter files (CLAUDE.md, settings.json, paper scaffold)
├── scripts/                 # CI helper scripts
├── .github/workflows/       # GitHub Actions (renders cheatsheet on push)
└── local_quarto_check.sh    # Local render validation

Limitations & Responsibilities

  • LLMs can hallucinate syntax or options. Always validate with quarto render + preview.
  • For clinical/PHI content, do not commit sensitive data. Use mock or de-identified examples.
  • If you change journals, update the CSL and reference-doc and re-render.