Research artifacts

Beyond scholarly papers, snowcite can ingest any user-supplied research material — interview transcripts, code, archival documents, notes, dataset descriptions — and cite it inline alongside the literature. This turns snowcite into a mixed-methods assistant rather than a pure literature-review tool.

Supported types

Type	What it is	Typical use
interview	Transcript of a conversation	Block quotes in findings/discussion
code	Source file or snippet	\lstinputlisting / #raw(read()) in methodology
document	Archival text, letter, internal doc	Cited like a primary source
note	Free-form research note	Referenced in discussion, background
dataset	Description of a dataset (not the data itself)	Methodology, data availability

Only text files are supported. Convert PDFs and .docx with pdftotext or pandoc first.

Importing

From a file:

Import my interview with P03:
  path = "interviews/p03.md"
  type = "interview"
  label = "P03 — first pilot"
  summary = "Pilot interview covering onboarding friction"
  metadata = {"participant": "P03", "consent": "written", "language": "ru"}

Claude calls import_artifact(path, type, label, summary, metadata).

From chat (for short notes / code snippets typed inline):

Save this snippet as a code artifact called "auth.py":
def login(user, pw):
    ...

Claude calls add_artifact_inline(type, label, content, ...).

Citation format

Every artifact gets a compact inline citation:

• [I:3] — interview id 3
• [C:auth.py] — code artifact (label, not id, is shown if available)
• [D:5] — document
• [N:2] — note
• [DS:1] — dataset

These tokens appear inline in the text, and the Primary-sources appendix expands them. Keep the format consistent throughout the document.

Assigning artifacts to a section

Outline entries accept artifact_ids alongside paper_ids:

save_outline([
  {
    "name": "findings",
    "target_words": 800,
    "paper_ids": [12, 17, 23],
    "artifact_ids": [1, 2, 3, 4],
  },
])

prepare_section_for_review and regenerate_section_brief load both automatically. The academic-reviewer subagent verifies quotes against both abstracts and artifact content.

Writing with artifacts

When Claude drafts a section that has assigned artifacts, it:

1. Reads the full content of each assigned artifact (not truncated like abstracts — you want verbatim quotes).
2. Weaves in quotes, paraphrases, references using the [X:id] citation labels.
3. Does not copy entire transcripts — just the relevant excerpts.

Example output:

Early users struggled to connect the two onboarding steps:

«Я не понял, надо ли сначала подтвердить почту или загрузить документ — инструкция была в двух разных местах.» [I:3]

This aligns with prior work on fragmented task flows [Smith2021].

Code inclusion

For code artifacts snowcite generates a backend-specific listing snippet:

Include the auth.py artifact in the methodology section.

Claude calls include_code_artifact(artifact_id, backend) and pastes the returned snippet into the section:

• LaTeX: \lstinputlisting[caption={auth.py}]{/path/to/auth.py}
• Typst: #figure(caption: [auth.py], raw(read("/path/to/auth.py"), lang: "python"))

Both read the file live from disk at compile time, so code changes propagate automatically.

Primary-sources appendix

Artifacts do not enter the main bibliography. Instead, generate an appendix:

Add a primary-sources appendix at the end of the document.

Claude calls generate_primary_sources_appendix(backend) and pastes the returned snippet after the main bibliography. Every included artifact appears with its citation token, label, type, and summary.

Excluding without deleting

set_artifact_included(artifact_id, False) removes an artifact from the writing pipeline (won't appear in prepare_section_for_review or the primary-sources appendix) but keeps it in the database. Useful for pilot interviews that don't make it into the final thesis.

set_artifact_included(artifact_id, True) puts it back.

For a hard delete, use delete_artifact(artifact_id).

What snowcite does not do

• No audio transcription. Bring transcripts ready.
• No PDF / .docx parsing. Convert with pdftotext / pandoc first.
• No data analysis. Use pandas / R / etc. externally, save figures as PNG, insert them as #image("plot.png") / \includegraphics.
• No automatic qualitative coding. You tag manually via metadata, or ask Claude to help interactively, but snowcite has no built-in codebook structure.