← ClaudeAtlas

synthesis-lablisted

Write grounded academic and application documents from a set of source papers, with strict citation discipline — every specific claim traces to a real supplied paper and citations are cross-checked against those papers before delivery, not fabricated. Use this skill WHENEVER the user wants to draft scholarly or application writing from sources: a literature review, theoretical framework, research proposal, introduction, discussion, conclusion, abstract, or academic essay — or a personal statement / statement of purpose (PhD, grad, study-abroad application) or résumé/CV. Trigger on phrasing like "help me write the lit review for ...", "draft the discussion section from these papers", "write my statement of purpose", "turn these sources into a research proposal", even when the tool isn't named. The defining value is citation grounding: no invented authors, years, or findings.
academicatstool-netizen/Cat_synthesis_lab · ★ 0 · AI & Automation · score 72
Install: claude install-skill academicatstool-netizen/Cat_synthesis_lab
# Synthesis Lab Turn a set of source papers plus the author's argument into a grounded academic document. Pure reasoning — no scripts. The entire point is **citation discipline**: every specific claim traces to a supplied paper, and you cross-check citations against those papers before delivering. One fabricated citation and an academic stops trusting the tool. ## Inputs you need - **The papers** — title, authors, year, and abstract/summary each. These can come from the `paper-search` skill, a deep-read report, or the user pasting them. If a paper's full text or deep-read claims are available, even better. - **The author's core argument** and **supporting points**. If blank, infer a draft from the papers + chosen output type and confirm with the user first. - **The output type** — see the catalogue in `references/synthesis.md`. ## How to run it **You are a single Claude playing every role below in sequence — there are no sub-agents and no separate model calls.** Run each step as its own reasoning pass in this one turn, carrying the result forward. The "plan / write / review / edit" roles are a discipline that keeps citations honest, not a multi-agent runtime. The planner's outline and any JSON are **internal scratch — do not show them to the user**; the only things you deliver are the final document, its references, and a one-line note on any citation you had to fix. **Default = write immediately. Don't gate behind a confirmation menu.** If you have the **document