Bestpaper Writing

Best-Paper Writing (v1.2): Abstract, Introduction, Method, Experiments, Figures

Turn a description of a research project (or a draft abstract) into an Abstract and Introduction that read like a recent CVPR/ICCV/NeurIPS award paper.

This skill is grounded in a sentence-by-sentence annotation of 55 award papers (2021–2025; 60 papers are analysed in total, and 5 of them (from 2024) are held out for testing, see PAPER_LIST.md in the repository root): what functions their abstracts and introductions contain, in which order, and how they read (references/corpus-profile.md). Award papers persuade through:

1. A sharp motivation — credit to specific prior work, then a concrete gap (what fails or is missing, when, and why), set up as a tension the paper resolves.
2. An idea about the problem — an observation from which the method follows, stated directly; a short handle for the idea when there is one (“next-brick prediction”).
3. Clear positioning — the closest work is named and one checkable difference is stated.
4. Confident, evidenced claims told as a story — results as claims with one or two headline numbers; the work is promoted where the evidence allows; a stance when the results support one. Weaknesses belong in the Limitations section, not in the Abstract or Introduction (author preference).
5. Plain, rhythmic prose. Each claim is stated once. Short, quotable sentences at the pivots (the bottleneck, the cause, the idea); the rest are full sentences (mean 20–24 words) that carry their reasons through because, while, by, which, not through trailing “, making …” clauses. Few logical connectors; colons allowed, no dashes (author preference). Mechanisms and examples instead of adjectives, short background, no implementation details.

Scope: title, Abstract and Introduction for all eight paper types (L1–L8, references/routing.md); also Method (references/method.md) and Experiments (references/experiments.md), mainly for method papers (L1/L2); and the figures of the paper (references/figures.md): a figure plan, Figure 1, the method overview and result figures, drawn where the content is known and left as planned placeholders where results are needed. Write the Abstract and Introduction first; Method, Experiments and figures follow the same claim map.

Workflow

Do not write prose before Step 4.

Step 0 — Intake (two input modes)

Mode A: intake form. If the user provides an intake JSON (exported from assets/intake_form.html, schema bestpaper-intake/v1), run python3 scripts/intake_to_brief.py <file.json> --out brief.md --bib refs.bib (paths relative to this skill folder) and work from brief.md; exit code 2 means a required field is missing. If the user describes the work in free text, load references/intake-form.md and map the description onto it. When a user has no material ready, point them to assets/intake_form.html (a local, offline form that exports JSON; the form is in Chinese).

Mode B: complete abstract in. If the user pastes an existing abstract (optionally with the introduction), do not rewrite it blindly. First reverse-engineer it into the intake fields of references/intake-form.md (task, dominant prior approach, gap and its reason, closest work, key insight, method components, results with numbers) and show the user a short table: field, what the draft says, status (present / vague / missing). Facts, numbers and terminology in the draft are the user's and are kept verbatim; never add a result, baseline or reference the draft does not contain. Then handle missing fields as below, and continue with Steps 1–8. In report.md add a change table (original sentence → revised sentence → reason, citing the award-paper pattern it follows) so the author can accept or reject each change.

In both modes:

• If a required field is missing (key idea, concrete gap with its reason, closest work, results with numbers), ask once in a compact list before drafting. In the same message, ask for the memorability fields that are empty (efficiency, integration with existing pipelines, supported inputs, applications, a concrete example, what is surprising, the trend it follows, the team’s name for the idea): these, not the prose, are what made award abstracts stand out in our comparisons. If the user cannot answer now, continue with [TODO: …] placeholders. Ask in plain text; do not depend on any agent-specific question tool.
• Never invent results, numbers, baselines, datasets or references.

Step 1 — Claim, tension, handle, stance, headline numbers

Write these for yourself (they go into report.md, not the paper):

• Claim: We show that [method/finding] [achieves what] [under which setting], because [key reason].
• Tension type: tradeoff / capability-missing / obvious-fix-too-costly / unexplained-phenomenon / open-question / belief-overturned (routing.md).
• Insight: the observation about the problem the method follows from, written as a plain statement (not “Our key insight is …”).
• Handle (optional): a short phrase naming the idea, distinct from the method’s name. Use the user’s term if they have one; otherwise propose one in the report. When there is a handle, put it in the title.
• Closest work: 1–2 works and the checkable difference (mechanism, scope, cost, assumption, input, output, evaluation).
• Stance (optional): what the work changes for the field, only if the results support more than “better on benchmark B”.
• Headline numbers: the 1–2 results that carry the claim.

Step 2 — Route

Load references/routing.md. Choose the paper type and motivation type; note the abstract arc and introduction plan.

Step 3 — Claim map

Load references/claim-calibration.md. List each claim with its evidence and the strongest wording the evidence supports. Use it to make claims confident where supported and scoped where not — never to add disclaimers. Mark headline / supporting / table-only numbers.

Step 4 — Skeleton

Load references/introduction.md and references/abstract.md. One line per paragraph (its message = its first sentence); one line per abstract sentence. Check against the award-paper tiers (introduction.md §2):

• always: credit to specific prior work → gap with reason → method → results;
• usual: opening setting or concrete scene; the insight just before the method;
• for method papers: a contrast with the closest work after the method;
• optional: contributions list, stance (at the end). The abstract defaults to Setting → Gap → Method → Result (abstract.md §3). Mark the pivot sentences (the bottleneck, the cause, the tension, the idea); they will be short and quotable. Then make a claim index: list every claim (gap, insight, each advantage, each result, stance) and the one paragraph where it appears. A claim listed under two paragraphs is merged into one. Generic field background gets at most two sentences.

Step 5 — Draft with exemplars

For each paragraph, read 2–3 award-paper paragraphs with the same function and, if possible, the same paper type: python3 scripts/find_exemplars.py --para <function> --type <L*>; for sentences, python3 scripts/find_exemplars.py <move> or --phrases <function>; for whole introductions, the route’s card in references/exemplars/cards/. Learn their order of information and rhythm; never copy a sentence or a run of ≥ 6 words. Write in the authors’ voice (“we”, “our model”); never describe the user’s input as “supplied”, “provided” or “reported”.

Step 5b — Method and Experiments (when requested)

1. Extend the claim map into an evidence map: claim → method subsection → experiment (table/figure) → ablation. Claims without evidence are softened in the introduction or get an experiment [TODO].
2. Method: follow references/method.md (opener, preliminaries, one subsection per component with need → mechanism → property, objective, optional discussion). Mechanisms, equations and losses the user has not given are [TODO]; never invent them.
3. Experiments: follow references/experiments.md (opener, setup, claim-first results per benchmark, ablations that mirror the claims, analysis, limitations). Tables use booktabs with \TODO{} cells for missing numbers.
4. Limitations belong in the experiments or a separate section; the author preference excludes them only from the abstract and introduction.

Step 5c — Figure plan and figures (whenever Method/Experiments are written, or when asked)

Load references/figures.md.

1. Make the figure plan from the evidence map: one row per figure with its place, the claim it carries, its type, panels, the material needed, its status and the caption's takeaway sentence (figures.md §2). Every main claim has a figure or a table; every figure carries one claim.
2. Choose Figure 1 by route (figures.md §3.2): diagnosis/contrast-results for failure, contrast-schematic for cost and insight, showcase for capability-gap/trend, plot for scaling and re-examination, concept for mechanism and theory, resource for datasets. In two-column templates it goes on page 1, usually full width.
3. Draw now what the user has stated: schematics of the mechanism (Figure 1 contrast panels, method overview, query or data-flow illustrations) in TikZ with assets/figures/figures_preamble.tex; parts whose mechanism is [TODO] are drawn dashed. Leave as placeholders everything that shows results (renders, qualitative grids, plots, numbers): a blank framed area of the planned size whose caption carries the full plan as \FigPlan{…}. Plots the user has data for use assets/figures/plot_style.py.
4. Captions open with the takeaway (figures.md §6); method colours, names and symbols are the same in every figure and match the text.

Step 6 — Output

• LaTeX template given (e.g., a CVPR author kit): load references/latex-output.md and follow it.
• No template: write Markdown (# Abstract, # Introduction).
• The paper text contains no move labels, workflow comments, or bracketed notes other than [TODO: …] / \todo{…}.

Step 7 — Check

1. Run python3 scripts/check_draft.py <files> on the abstract and the introduction. Fix every ERROR. Each WARN means the draft is outside the range of 90% of award papers (connectors, numbers, hedges, boosters, “key insight” formulas, sentence length, runs of short sentences, trailing participial clauses) or shows a known AI-draft habit — revise or justify it in the report. NOTEs are information (e.g., the draft promotes the work less than award papers do). python3 scripts/style_metrics.py <files> shows every metric with its corpus percentile.
2. If the paper has figures, run python3 scripts/check_figures.py main.tex: fix every ERROR; each WARN (caption length, long first sentence, raster plots, absolute widths, too many floats) is fixed or justified; NOTEs list the placeholders still to fill.
3. Semantic checks:
   • Specific prior work is credited before the gap; the gap is concrete (what fails, when, why) and the tension is resolved by the idea.
   • The insight is an observation about the problem, stated directly, not a hypothesis and not a description of the architecture.
   • The closest work is named with one checkable difference (method papers).
   • Results are claims at the strongest supported level, with headline numbers; the work is promoted where evidence allows; nothing is claimed that the evidence does not support (no unscoped “first”, “state-of-the-art on all benchmarks”).
   • No limitation, failure case or disclaimer in the Abstract or Introduction; mixed results are scoped to the wins.
   • No defensive qualifiers about the protocol; no implementation details.
   • Abstract, introduction and contributions state the same claims with the same strength.
   • Nothing is said twice inside the introduction (insight, research questions, results, advantages); the contributions do not repeat the results paragraph.
   • Read the draft aloud once: if several consecutive sentences each carry one bare fact, merge them so that each sentence carries a relation (cause, contrast, means, consequence).

Step 8 — Deliver

1. The paper files (LaTeX project or Markdown).
2. report.md for the authors: claim, tension type, insight, handle, closest work, route, claim map, check_draft summary, [TODO] list, questions, notes for the Limitations section (weaknesses deliberately left out of the abstract/introduction), and, when figures were planned, the figure plan table with the material each placeholder needs and the check_figures summary.

Hard rules

• Never fabricate numbers, datasets, baselines, citations or results.
• Never cite a key that is not in the user’s bibliography; mark missing references with [TODO: cite …].
• No limitations, self-undermining sentences or defensive qualifiers in the Abstract or Introduction.
• The Abstract and Introduction tell the story; tables tell the numbers.
• No generic openings (“has attracted increasing attention”, bare “plays an important role”) and no adjectives about the method without a concrete follow-up.
• Do not frame the work as a patch on a naive baseline (ruling out an obvious alternative with a quantified reason is fine).
• Keep terminology fixed once introduced.
• No dashes in the title, abstract or introduction (author preference): no em dash (—, ---) and no spaced en dash. Rewrite with a colon, a comma, a short second sentence or parentheses. Numeric ranges (2–3) and hyphenated compounds (image–motion, train–test) are fine. check_draft.py reports each dash as an ERROR. Colons are allowed.
• Name mechanisms, not abstractions. When the input does not say how a component works (who inspects, what is compared, how something moves), write [TODO: …] and ask; do not fill the gap with abstract nouns (“state”, “record”, “inspection”, “protocol”) that sound complete but say nothing.
• Describe components by their relation, not as parallel items: what problem each one answers and what the next one fixes (“A keeps …; because A alone lets errors persist across edits, B …”), rather than “A does … . B does … . C does … .”
• Figure and table numbers from notes or intake forms refer to the author’s full draft. Do not cite them in the abstract or introduction unless the figure is part of the paper being written; list the example and its figure in report.md.
• One aphorism. Use at most one or two contrastive maxims (“X is not Y”) for the thesis; state the rest as plain fact.
• Keep what makes the draft memorable. When revising for a checker WARN, do not merge away the short pivot sentences, the cause sentence, or the handle in the title; fix the WARN elsewhere or justify it. A memorable draft that is slightly off the corpus statistics beats a statistically average one.
• Never fabricate figure content. No invented result images, renders, curves or numbers in figures; schematics show only mechanisms the user has stated. Results figures stay placeholders with their plan in the caption until the authors supply the material.
• Do not end sentences with a participial clause that restates significance (“…, making X possible”, “…, extending Y”, “…, bringing Z”). Award papers do this in ~8% of sentences; AI drafts in 30% or more.

Files

Use python3. Attribution and licences: NOTICE.md.