Scaffold a literature review.
The map of the field before you write the prose — where each paper fits, what the disagreements are, and where the gap you're actually filling lives.
You have a reading list. You need to write a lit review section. You could start by reading each paper and writing about it in order — that produces a chronology, not a lit review. This prompt gets you the structural map first: what the debates are, where each paper sits, and where your own contribution lives.
THE PROMPT
I'm scaffolding a literature review. Attached are [N] papers. Give me the structural map before I write, in this format: **The three main threads in this literature** - One thread per bullet. For each: name the debate or research direction, and note which of the attached papers contribute to it. **Where each paper sits** - Bulleted list, one per paper. Format: **[Author (Year)]** — one-sentence summary of contribution — one-sentence note on what thread it belongs to and how it disagrees with adjacent papers. **The disagreements worth flagging** - 2-4 bullets. Specific unresolved disputes in this literature. Cite the papers on each side of each dispute. **The gap I'm plausibly filling** - 1-2 sentences. Given the map above and my stated research question ([QUESTION]), where does my contribution actually sit? Be honest — if my question has been answered already by one of these papers, tell me. **Prose scaffold** - A three-paragraph outline for the lit review section itself. First paragraph frames the field; second maps the debates; third positions my work. One sentence per paragraph, describing what it does structurally. Papers attached: [ATTACH PDFs] My research question: [DESCRIBE YOUR QUESTION IN 1-2 SENTENCES]
**The three main threads in this literature**
- Scaling laws and their extrapolation limits — Kaplan (2020), Hoffmann (2022), Muennighoff (2023).
- Emergent capabilities as artifact vs real phenomenon — Wei (2022), Schaeffer (2023).
- Compute-optimal training for downstream task performance — Hoffmann (2022), Villalobos (2023).
**Where each paper sits**
- **Kaplan et al. (2020)** — Establishes power-law relationships between model size, dataset size, and loss — belongs to the scaling-laws thread, and is directly disagreed with by Hoffmann on the compute-optimal question.
- **Hoffmann et al. (2022, Chinchilla)** — Argues Kaplan-style scaling systematically undertrained models relative to data — bridges scaling-laws and compute-optimal threads.
- **Wei et al. (2022)** — Documents emergent capabilities appearing discontinuously at scale — foundational for the emergence thread.
- **Schaeffer et al. (2023)** — Argues emergence is an artifact of discontinuous metrics; smooth metrics would show gradual improvement — directly disagrees with Wei.
- [... continues per paper ...]
**The disagreements worth flagging**
- Kaplan vs Hoffmann on compute-optimal — both empirical, differ in what they held constant. Modern practice sides with Hoffmann.
- Wei vs Schaeffer on emergence — this is a live debate as of 2026, not resolved. Any claim your paper makes about emergence needs to engage with both sides.
**The gap I'm plausibly filling**
Given your stated question — [QUESTION] — you sit at the intersection of the emergence debate and the compute-optimal thread, but neither paper set has addressed the specific angle you're proposing. Genuine contribution, not a duplicate.
**Prose scaffold**
- Paragraph 1: Introduces the scaling-laws project as the frame for modern LLM research, from Kaplan through Hoffmann.
- Paragraph 2: Maps the emergence debate, noting Wei's original claim and Schaeffer's critique, and situates the compute-optimal literature as adjacent but under-connected to it.
- Paragraph 3: Names the specific under-explored question your paper addresses, and how it draws on both threads to answer it.