AI Coding Tools for CLIF
Multiple AI-assisted coding setups are available for CLIF development. Choose the one that fits your workflow.
Table of contents
- Overview
- Comparison
- Option 1: Official CLIF Skills (Marketplace)
- Option 2: Kaveri’s Claude Setup
- Option 3: J.C.’s clif_skills_agents
- Quick Config Files
- Which Should I Use?
- Contributing
Overview
The CLIF community has developed several AI coding assistant configurations. All include CLIF-specific knowledge (schema, mCIDE vocabularies, best practices) to help you write correct code faster.
Comparison
| Official CLIF Skills | Kaveri’s claude-setup | J.C.’s clif_skills_agents | |
|---|---|---|---|
| Repo | Common-Longitudinal-ICU-data-Format/skills | kaveriC/claude-setup | RushAI-jcr/clif_skills_agents |
| Install | Marketplace (/plugin install) | Copy files + marketplace skill | Copy files (standalone) |
| Platforms | Claude Code | Claude Code | Claude Code, Kilo Code (VS Code), Codex CLI |
| Languages | Python | Python | Python and R |
| Notebooks | — | — | marimo, Jupyter, Quarto, RMarkdown |
| CLIF Schema | Full mCIDE + schemas | References official skill | Full 2.1 schema embedded |
| Session Commands | — | /start-clif-session, /end-clif-session | /start-session, /end-session |
| Best For | Marketplace users | Quick Claude Code setup | Multi-platform teams, R users |
Option 1: Official CLIF Skills (Marketplace)
The canonical, consortium-maintained skill. Install via Claude Code’s plugin marketplace.
Installation
# Inside Claude Code
/plugin marketplace add Common-Longitudinal-ICU-data-Format/skills
/plugin install clif-icu@clif-skills
What’s Included
- Full CLIF 2.1 data dictionary
- mCIDE controlled vocabularies (labs, vitals, meds, etc.)
- YAML schema definitions
- clifpy API reference
- Example scripts
This is the recommended approach for most users. The skill is maintained by the consortium and stays up-to-date with CLIF schema changes.
Option 2: Kaveri’s Claude Setup
A quick-start configuration that installs the official skill plus adds useful workflow features.
Installation
# Clone the repo
git clone https://github.com/kaveriC/claude-setup
# Copy global config
cp claude-setup/CLAUDE.md ~/.claude/CLAUDE.md
cp claude-setup/settings.json ~/.claude/settings.json
cp claude-setup/commands/*.md ~/.claude/commands/
# Install CLIF skills plugin (inside Claude Code)
/plugin marketplace add Common-Longitudinal-ICU-data-Format/skills
/plugin install clif-icu@clif-skills
What’s Included
| File | Purpose |
|---|---|
CLAUDE.md | Workflow rules: plan mode, verification, task tracking, lessons learned |
settings.json | Extended thinking enabled, CLIF plugin, voice notification hook |
commands/start-clif-session.md | /start-clif-session — reads progress, presents state |
commands/end-clif-session.md | /end-clif-session — documents progress, saves lessons |
Best For
Claude Code users who want:
- Structured session management
- Voice notifications when Claude finishes
- Built-in task tracking and lessons learned
Option 3: J.C.’s clif_skills_agents
A comprehensive, multi-platform setup supporting Python, R, and multiple notebook types.
Installation
Claude Code (recommended):
cd your-clif-project/
cp -r /path/to/clif_skills_agents/.claude .claude
cp /path/to/clif_skills_agents/CLAUDE.md ./CLAUDE.md
cp /path/to/clif_skills_agents/SKILL.md ./SKILL.md
Kilo Code (VS Code):
cp /path/to/clif_skills_agents/AGENTS.md ./AGENTS.md
cp -r /path/to/clif_skills_agents/.kilocode .kilocode
Codex CLI:
cp /path/to/clif_skills_agents/AGENTS.md ./AGENTS.md
Auto-Detection
Just run /start-session — the assistant auto-detects your setup:
| Your project files | Language | Notebook | Skills loaded |
|---|---|---|---|
.py + marimo in pyproject.toml | Python | marimo | clif-data, python-dev, marimo, batch, figures |
.ipynb only | Python | Jupyter | clif-data, python-dev, jupyter, figures |
.R / .Rmd / renv.lock | R | RMarkdown | clif-data, r-dev, r-data-tooling, rmarkdown, figures |
.qmd | R | Quarto | clif-data, r-dev, r-data-tooling, quarto, figures |
8 Rules That Prevent Cross-Site Bugs
The assistant enforces critical rules automatically:
- Filter on
_category, never_name— Categories are the CLIF controlled vocabulary -
Validate before analysis — Python: clif.validate_all()R: CLIF-TableOne - Use ADT for ICU times, not hospitalization — Filter
adtwherelocation_category == 'icu' - Log row counts at every filter step — CONSORT-style attrition tracking
- Use encounter stitching — One ICU stay can span multiple hospitalization records
- Distinguish continuous vs intermittent meds — Drips vs boluses
- All timestamps are UTC — Timezone conversion is mandatory
- JAMA figure style — Arial, 8pt min, no overlap, legends outside plot area
Best For
- R users — Only option with R support
- VS Code users — Works with Kilo Code extension
- Multi-notebook teams — Supports marimo, Jupyter, Quarto, RMarkdown
- Standalone setups — No marketplace dependency
Quick Config Files
Python (clif_config.json)
{
"data_directory": "/path/to/clif/parquet/files",
"timezone": "US/Eastern",
"site_name": "YOUR_SITE"
}
R (config/config_template.yaml)
site: "YOUR_SITE"
file_type: "parquet"
clif_data_path: "/path/to/clif/parquet/files"
site_time_zone: "US/Eastern"
Which Should I Use?
Start with the Official CLIF Skills if you:
- Use Claude Code
- Want the consortium-maintained, always-up-to-date skill
- Prefer marketplace installation
Add Kaveri’s setup if you also want:
- Session management commands
- Voice notifications
- Task tracking built-in
Use J.C.’s clif_skills_agents if you:
- Need R support
- Use VS Code with Kilo Code
- Want a self-contained setup without marketplace
- Work with multiple notebook types
These are complementary! Kaveri’s setup actually installs the official skill. J.C.’s is a standalone alternative for broader platform coverage.
Contributing
Have improvements to share? PRs welcome to any of these repos: