Getting started

1. Create a project directory

mkdir my-review && cd my-review

That's it — .snowcite/ and every generated file land here. cd into this directory any time you want to work on this review; snowcite resolves the active project from your current working directory.

2. Initialize

From Claude Code (inside my-review/):

Let's start a snowcite project. I'm writing a bachelor thesis on adversarial attacks on LLMs, in Russian, ГОСТ 7.32, Typst.

Claude collects metadata via AskUserQuestion (author, institution, deadline, etc.), then calls init_project(metadata=...). The tool:

• Creates .snowcite/papers.db and .snowcite/cache/.
• Writes a tailored CLAUDE.md at the project root — this guides every future Claude session in this directory.
• Writes .claude/agents/academic-reviewer.md and .claude/agents/humanizer.md.
• Emits a diff for .claude/settings.json (new file or missing entries) so Claude can ask you how to resolve it: merge, overwrite with backup, or skip.

3. Set review criteria

Criteria are free text — include clauses, exclude clauses, and optionally user-defined clusters for the review summary:

Include: adversarial attacks on language models (LLMs or text classifiers),
         papers from 2023+.
Exclude: vision-only adversarial work, hardware attacks.
Clusters: attacks-optimization, attacks-jailbreak, defenses-detection,
         defenses-certified, evaluation, surveys.

Claude calls set_review_criteria(criteria_text=...).

4. Search and review

Search arXiv, Semantic Scholar and OpenAlex for "adversarial attacks LLM
2023-2025", limit 50 each.

Claude calls search_papers(query, limit=50). By default, auto_save=True persists results to the DB and returns {saved, duplicates, new_ids, titles} — abstracts never enter your chat context. Then:

Start reviewing. Batch of 20.

For each paper in the batch, Claude decides autonomously and tags the decision with reviewed_by (auto_high for clear matches, auto_low for less certain, or defers to you for genuinely borderline ones). After each batch it saves a rolling review summary.

5. Snowball

Once you have an approved set, deepen through citations:

Run snowball — expand_citations(references) on the approved papers, then
re-review the new ones.

New papers from the snowball are added to unreviewed; the summary is marked stale. Iterate.

6. Outline and skeleton

Propose an outline for the document.

Claude calls save_outline(sections=[...]) with sections, paper IDs per section, and target word counts. You approve or edit. Same for save_skeleton(sections=[...]) — the 3-5-sentence version of each section — which gives you the arc in about 500 words total.

7. Expand section by section

Write the first section.

Claude drafts, calls check_section_drift, then save_section. It then spawns the academic-reviewer subagent, which reads your section + the assigned papers' abstracts cold (fresh context) and returns a structured findings list — unsupported claims, citation misuse, fabricated quotes, logical gaps. Decide which to fix; save the revision.

8. Polish

After all sections are in:

Run polish_document for cross-section consistency, then humanizer.

polish_document handles transitions and cross-section duplication. Then the humanizer subagent flags machine-translated phrasing, LLM tics and awkward sentences — per-phrase replacement suggestions you accept or reject.

9. Compile

compile_pdf on review.typ

typst compile runs and drops review.pdf next to the source. Done.