Skip to content

Workflow

snowcite is five phases. Each phase persists state to the project DB, so /clear or a new session doesn't lose work — get_session_state() returns a snapshot telling Claude where it left off.

flowchart TD
    A([Onboarding]) --> B([Search & review])
    B --> C([Structure: outline → skeleton])
    C --> D([Write per section + academic-reviewer])
    D --> E([Polish + humanizer → compile])

Phase 1 — Onboarding

One MCP call: init_project(metadata=...). See Getting started.

Re-run with update=True when metadata changes mid-project. Re-run with update_agents=True to pull the latest subagent prompts without losing your project state.

Phase 2 — Search and review

Every new project starts with set_review_criteria. Before each review batch, Claude calls get_review_criteria() as a drift guard (criteria drift is common when reviews span weeks).

search_papers(query, sources=None, limit=20, auto_save=True, abstract_max_chars=0)

  • sources=None → automatic routing based on your discipline metadata. STEM (cs, physics, math) includes arXiv; medicine/biology includes PubMed; everyone gets OpenAlex + Semantic Scholar + Crossref.
  • auto_save=True → results go straight into the DB. The response is {saved, duplicates, new_ids, titles} — no abstract bodies in the chat.

Review batches

get_unreviewed_papers(limit=20, include_abstracts=False) returns compact metadata (title, year, venue, authors) — enough for clear-cut classifications without leaking harmful-sounding abstracts into the context (which accumulates and can trigger safety refusals on sensitive topics like adversarial ML).

For each paper Claude decides:

  • Clear match → batch set_review_status([...], "approved", reviewed_by="auto_high")
  • Clear miss → batch set_review_status([...], "rejected", reviewed_by="auto_high")
  • Likely but uncertain → decide with reviewed_by="auto_low" — the user can later pass over just these via get_low_confidence_reviews()
  • Genuinely borderline → defer. For borderline papers, fetch the full abstract via get_paper_details(paper_id), summarize in 1-2 neutral sentences, and show to the user. Do not recommend a decision — that creates bias.

After each batch, save_review_summary(summary, clusters) updates the rolling ≤500-word summary.

Snowball

expand_citations(paper_id, "references" | "citations") walks the citation graph of an approved paper. Uses Semantic Scholar as the graph source; for arXiv-only papers without a DOI, the call falls back to DOI lookup. New papers auto-save to unreviewed; the summary is marked stale.

Phase 3 — Structure

save_outline(sections=[{"name", "target_words", "paper_ids": [...]}, ...]), then approve_outline() after the user OKs it.

save_skeleton(sections=[{"name", "draft"}, ...]) — 3-5 sentences per section. This gives you the document arc in ~500 words total. Approve via approve_skeleton().

Approval is semantic, not enforced: tools keep working if you skip it, but the workflow's quality guarantees (drift checks against the approved outline, subagent context) assume it.

Phase 4 — Section by section

For each outline entry:

  1. Claude drafts the section based on the skeleton, assigned paper abstracts, and any previously expanded sections.
  2. check_section_drift(name, content) — returns warnings if the word count exceeds max(100, ±30%) of target, or paper IDs have diverged from the outline. Claude surfaces these to the user before saving.
  3. save_section(name, content) persists with a version counter.
  4. Claude spawns the academic-reviewer subagent via the Agent tool. It calls prepare_section_for_review(name) and returns a structured findings list.
  5. User picks fixes; Claude applies them via another save_section.

Phase 5 — Finalize

polish_document([...]) — Claude rewrites to fix cross-section transitions, consistent terminology across sections, and removes duplicate tezises. This is structural, not stylistic.

Then spawn the humanizer subagent. It flags language issues (machine-translated phrasing, LLM tics, awkward word choice) and proposes per-phrase replacements. User accepts; polish_section(name, polished_content) persists with polished=1.

Finally compile_pdf(doc_path). Backend is inferred from the extension:

  • .typ → typst compile
  • .tex → tectonic

Session recovery

After /clear or in a new session, Claude calls get_session_state() first. It returns the current phase (reviewing / writing / etc.) and a next_action hint, so Claude can resume without re-probing the DB.