Academic Paper Explainer in Plain Language

Prompt

Act as a science writer who has sat with researchers and still writes for a smart non-specialist. Explain only what is in Inputs. If the PDF is not pasted, say what you cannot know. Do not invent citations, quotes, p-values, or sample sizes.

Inputs:
- Paper ID: [Title, authors if known, year, venue if known]
- Pasted text: [Abstract required; methods/results if you have them]
- Audience: [PM / reporter / undergrad / exec]
- Why I care: [Decision or curiosity]
- Jargon I already know: [List]
- Must avoid: [Medical advice, investment advice]
- Length: [Short brief / medium]

Generate:
1. Confidence label: What you actually had (abstract only vs methods vs results). What is therefore GUESS or unknown.
2. Plain-language claim: 5-8 sentences. One sentence on what the paper is not claiming, if that is clear from the paste.
3. Who and how: Population, method family, comparison, in words from the paste. If missing, write unknown, need methods.
4. Results that are in the paste: Numbers only if they appear. No invented effect sizes.
5. Limits: From the paper's own caveats if pasted; otherwise typical limits labeled as general, not as this paper's.
6. Glossary: 6 terms, each 1 sentence, no new jargon in the definition.
7. So what for Why I care: 6 sentences. A misuse (overclaim) to refuse.
8. Cite-as: A citation line using only IDs in Inputs. If a DOI is missing, do not invent one. Further reading: none unless the user pasted names.

Constraints:
- No em dashes. No fake quotes from authors.
- No medical, legal, or investment advice.
- If Pasted text is empty, stop and ask for an abstract. (If you still illustrate, mark it as a template, not as that paper.)
- Do not add papers to a reference list.
- Write at Audience level. Short sentences.

Instructions

Replace every [bracket] with your details before running. Works on ChatGPT, Claude, and Gemini.

Examples

Example Input

Paper ID: Attention Is All You Need, Vaswani et al., 2017, NIPS
Pasted text: ABSTRACT ONLY. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.
Audience: product manager who is not an ML researcher
Why I care: we keep hearing "transformer" in vendor decks; I need to know what the original paper actually claimed so I do not buy snake oil
Jargon I already know: GPU, training, model
Must avoid: investment advice, medical
Length: medium

Example Output

1. Confidence
Abstract only. No methods section, no tables, no authors' quotes beyond this abstract. BLEU numbers below are in the paste. "Why it works" details are unknown without the paper body. Do not treat this as a full read.

2. Plain-language claim
The authors propose a model called the Transformer that uses attention only. It does not use recurrence or convolutions. On two translation tests in the abstract, they report higher BLEU and less training time, with more parallelization. They also say it worked on English constituency parsing, including with limited data. This abstract does not claim that Transformers are a general intelligence, a search engine, or your vendor's product.

3. Who and how
Tasks named: WMT 2014 English-to-German and English-to-French translation, plus English constituency parsing. Method family: attention-only neural net. Comparisons: "existing best results, including ensembles" and "best models from the literature" (names not in this abstract). Hardware mentioned: 3.5 days on eight GPUs for the French result. Dataset sizes and full method: unknown, need methods.

4. Results in the paste
28.4 BLEU on WMT 2014 En-De, "over 2 BLEU" above the then-best including ensembles. 41.8 BLEU on WMT 2014 En-Fr, called a single-model state-of-the-art in this abstract, after 3.5 days on eight GPUs. Parsing: "successfully," no number in the paste. No p-values in the paste; do not invent them.

5. Limits
From this abstract: claims are about translation quality (BLEU) and train time on those tasks, plus a parsing transfer. Not in the paste: fairness, hallucination, your domain, cost in 2026 hardware. General (not this paper's own words): BLEU is not "understanding."

6. Glossary
Attention: the model weights which other words to use when building a word's representation.
Transformer: the architecture in this paper that uses attention instead of recurrence or convolution.
Recurrence: stepping through a sequence in order, which is harder to parallelize.
Convolution: a local filter, more common then in vision-style nets.
BLEU: an automatic overlap score for translation, not a human rater.
Parallelizable: more of the work can run at once on GPUs.

7. So what
"Transformer" in a deck should mean this kind of architecture, not a miracle. The original claim is translation quality and train time on named benchmarks. Misuse to refuse: "the 2017 paper proves their chatbot is accurate on our bank data." It does not. Investment advice: none.

8. Cite-as
Vaswani et al., 2017, Attention Is All You Need, NIPS. DOI: not in Inputs, not invented. No further reading list.