export-report - SKILL.md Agent Skill

name: export-report description: Compile a structured research report from a vitrine study — gathers cards, decisions, annotations, and scripts into an organized REPORT.md. tier: community category: system

Export Research Report

Compile all artifacts from a vitrine study — cards, decisions, annotations, scripts, protocol, and results — into a single structured REPORT.md.

When to Use This Skill

User asks to export, compile, or summarize a study
User wants a readable report from a completed research session
User wants to share findings with collaborators who don't have vitrine access
Before archiving or publishing study results

Setup

from vitrine import study_context, register_output_dir, show, list_annotations
import pandas as pd
from pathlib import Path

STUDY = "<study-label>"  # provided as context when dispatched
output_dir = register_output_dir(study=STUDY)

Procedure

Step 1: Gather Context

Call study_context(STUDY) and capture:

cards — all cards with card_id, title, type, preview
decisions_made — all decisions with outcomes
pending_responses — any unresolved blocking cards

Call list_annotations(study=STUDY) to get researcher notes. Each annotation has id, text, timestamp, card_id, card_title.

Step 2: Read Existing Artifacts

Read these files from output_dir if they exist:

PROTOCOL.md — study protocol (research question, design, analysis plan)
RESULTS.md — study results and conclusions

List all .py files in output_dir / "scripts/". For each script, read the file to extract its docstring (first triple-quoted string) as the script description. If no docstring, use the filename.

Step 3: Categorize Cards

Group cards by type and role:

Study description: The first markdown card (usually titled "Study Description") or content from PROTOCOL.md
Data cards: Cards with type TABLE — these are results tables
Plot cards: Cards with type PLOTLY or IMAGE — these are figures
Decision cards: Cards with type DECISION — these are researcher choices
Finding cards: Other markdown cards — these describe methods, findings, limitations

Step 4: Compile REPORT.md

Build the report with these sections:

# Research Report: <Study Title>

**Study:** <study-label>
**Generated:** <ISO date>
**Cards:** N | **Scripts:** N | **Decisions:** N

---

## 1. Study Description

[Content from the first markdown card or PROTOCOL.md. Include the research
question, study design, population, variables, and analysis plan.]

## 2. Methods

### Scripts

| # | Script | Description |
|---|--------|-------------|
| 1 | `01_cohort_definition.py` | Define sepsis cohort from MIMIC-IV |
| 2 | `02_baseline_characteristics.py` | Compute baseline demographics |
| ... | ... | ... |

### Protocol

[Full content of PROTOCOL.md if it exists, or a note that no protocol was saved.]

## 3. Results

### Key Findings

[Summarize finding cards — markdown cards that describe results, effect sizes,
confidence intervals, and interpretations.]

### Tables

[For each TABLE card, include the card title, description, row/column count,
and a preview of the first few rows formatted as a markdown table. Include the
source field if present.]

### Figures

[For each PLOTLY or IMAGE card, note the card title and description. Since
markdown cannot embed interactive plots, describe what each figure shows and
reference the plot file in `plots/` if available.]

## 4. Decisions Log

| # | Decision | Choice | Rationale |
|---|----------|--------|-----------|
| 1 | Severity score? | SOFA | Standard for Sepsis-3 studies |
| 2 | Exclude readmissions? | Yes | One observation per patient |
| ... | ... | ... | ... |

[For each decision card, extract the question asked, the choice made (from
response action or form values), and any message the researcher provided as
rationale. Include timestamps.]

## 5. Scripts

[For each script, include the full source code in a fenced code block with
the filename as heading. This makes the report fully self-contained.]

### 01_cohort_definition.py

```python
<full script content>
```

### 02_baseline_characteristics.py

```python
<full script content>
```

## 6. Annotations

| Card | Annotation | Timestamp |
|------|------------|-----------|
| Cohort Preview | "Looks good, but check age range" | 2026-02-10T14:30:00 |
| ... | ... | ... |

[Include all researcher annotations. If there are none, note that no
annotations were recorded.]

---

*Report compiled from vitrine study `<study-label>`.*

Step 5: Save REPORT.md

Write the compiled report to output_dir / "REPORT.md".

Step 6: Post Confirmation Card

Post a single card confirming completion with a brief table of contents:

toc = f"""## Report Exported

`REPORT.md` saved to study output directory.

**Contents:**
1. Study Description
2. Methods ({script_count} scripts)
3. Results ({table_count} tables, {plot_count} figures)
4. Decisions Log ({decision_count} decisions)
5. Scripts (full source)
6. Annotations ({annotation_count} notes)

**Total cards processed:** {card_count}
"""

show(toc, title="Report Exported", study=STUDY)

Critical Implementation Notes

Read every script file. The report includes full script source so it is self-contained and reproducible without access to the original output directory.
Decision extraction. Decision cards store responses differently depending on whether they used forms or quick actions. Check both response_action and form_values in the card data. The study_context() return includes decisions_made which has already-resolved decisions.
Table previews. For TABLE cards, the preview field in the card dict contains inline data. Format it as a markdown table, limiting to 10 rows max. If the preview is truncated, note the total row count.
Plot descriptions. Interactive Plotly figures cannot be embedded in markdown. Instead, describe what the figure shows using its title and description fields. Reference the corresponding PNG file in plots/ if one exists.
Annotation ordering. list_annotations() returns newest first. Reverse the order in the report so annotations read chronologically.
Missing artifacts are normal. Not every study will have a PROTOCOL.md, RESULTS.md, annotations, or decisions. Handle each gracefully with a note like "No protocol was saved for this study."
Use study=STUDY on every show() call. This keeps the confirmation card grouped with the study.