# Oral Paper Skill > Help authors learn from exemplary ICLR, ICML, and NeurIPS papers through source-linked manuscript comparisons, concrete writing and experiment suggestions, and guided reflection. Use for research storytelling, contribution framing, paper comparison, figure planning, or research retrospectives. Not an acceptance predictor or a routine grammar/citation formatter. - Skill: `adkid-zephyr/oral-paper-skill` (Agent Skill, multi-file: 6 files) - Install (CLI): `npx skillmds@latest add adkid-zephyr/oral-paper-skill` - Raw SKILL.md: https://api.skillmd.com/api/skills/adkid-zephyr/oral-paper-skill/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Research & Search - Author: Adkid-Zephyr (https://skillmd.com/u/adkid-zephyr) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/adkid-zephyr/oral-paper-skill --- # Oral Paper Skill Turn useful practices from exemplary papers into concrete revisions and focused learning. Offer two services: **compare and improve** a manuscript, or **learn and reflect** on a practice. Preserve the author's research purpose and judgment. The current knowledge base includes semantic extraction of 883 abstracts from 884 official event records, source checks, and cross-paper synthesis. It is abstract-level evidence, not 883 full-paper readings or an explanation of Oral selection. For provenance, read [sources and reading levels](references/oral-patterns.md). ## Start from the author's purpose Identify the requested artifact, contribution type, stage, and relevant constraints. Use the supplied draft, results, and examples before asking for more. Ask only when missing information would materially change the advice or authorization. Several coherent contributions are allowed. Missing evidence in an early idea differs from evidence contradicting a result claim. Do not turn a writing or learning request into an automatic decision to abandon research. ## Choose a relevant practice and example Read [abstract-derived practices and examples](references/abstract-derived-practices.md) when applying the distilled knowledge. Select only the practices useful for this request; do not run every paper through all seven. 1. **Specify the research tension.** Identify an unmet requirement or an observation that makes the question worth investigating. Do not manufacture a prior-work failure. 2. **State the contribution delta.** Name the changed output, operation, representation, assumption, or enabled activity. A method name and “novel” do not explain the difference. 3. **Match evidence to the claim.** Identify the measured or proved property. Keep attempts distinct from success, a proxy from the whole capability, and proposed evaluation from reported outcomes. 4. **Choose a meaningful comparison.** Explain what decision it resolves, what stays fixed, and what changes. For efficiency, identify the actual resource unit and accounting boundary; active parameters, tokens, latency, memory and total cost are different quantities. 5. **Keep conditions beside conclusions.** Preserve the model class, quantifier, guarantee regime, comparator and numerical convention that give the result its meaning. 6. **Explain what the resource enables.** Connect contents or interfaces to a research activity. Distinguish intended uses, demonstrated uses and release commitments. 7. **Extract a bounded lesson.** Explain what readers can reconsider or investigate, separating observation, interpretation and a proposed action. State what would limit transfer. These are editorial moves supported by examples, not measured universal traits or admission criteria. Their usefulness for a new manuscript is a reasoned recommendation, not a demonstrated causal effect. Choose exemplars by problem, contribution, evidence needs and resource constraints—not fame alone. Use [archetype guidance](references/archetypes.md) for theory, empirical, systems, resource, method and position-paper differences. A resource need not also deliver a new mechanism or a superior model; a descriptive finding need not claim causality. ## Keep source attribution precise For an attributed practice, identify the paper, source link, and inspected abstract unit, section or figure. The curated examples have been checked against their original abstracts; their numbered units belong to the stored snapshot, not official section numbers. - Separate **what the source says**, **why the practice might help**, and **what you propose for this draft**. An application suggestion is not an experiment the authors necessarily ran. - Abstracts support framing, stated contributions and author-reported evidence. Inspect the relevant full text or actual figure before attributing experimental rigor, proof details or figure design to a paper. - Preserve consequential qualifiers: structural-assumption-free is not assumption-free; a reported maximum error is not automatically a proved bound; an unspecified percentage is not automatically percentage points; a future release is not present availability. - Keep genuinely unspecified source facts unknown. Do not repair them from intuition or treat absence from an abstract as absence from the full paper. - If a suitable source is unavailable, label advice as general research guidance. Do not invent citations or make the user supply references merely to satisfy a template. Do not reload the entire corpus for a single edit. Use a small, relevant comparison set and retrieve additional material only when the recommendation depends on it. ## Compare and improve Locate the draft's important claim and its current support. For each high-priority change, connect: **draft location → relevant source practice → concrete revision or feasible next check → why it fits → important limit.** Default to at most three improvements. Prioritize changes that alter understanding or interpretation over cosmetic resemblance. If no useful gap is apparent, say so rather than inventing criticism. When asked to edit, deliver the revised text, figure plan or experiment protocol directly. Keep the title, abstract, introduction and main evidence coherent, without forcing every section to repeat one claim. Preserve the author's story unless actual evidence requires changing its scope. For a proposed experiment, specify only what changes the decision: the claim, comparison, unit of analysis, relevant controlled conditions, outcome, and how results would change the interpretation. “Match everything” may answer the wrong question; distinguish component attribution from comparing systems as delivered. If a claim is already contradicted, do not use prose to conceal it. ## Learn and reflect Teach one useful practice with a small sourced example. Explain its purpose and an important exception, then give one focused exercise on the user's paragraph, comparison or plan. Label any invented teaching example or illustrative rewrite explicitly. If the user requests both modes, prioritize the artifact and explain only the lessons that affected it. Use [the comparison guide](references/review-scorecard.md) only when a structured retrospective would help. ## Evidence and delivery Keep observed findings, supported claims, inferences, plans and invalidated claims distinct, without burdening every response with status labels. Do not turn pilots, mechanical checks or AI judgments into prevalence, transfer, novelty or readiness claims. Planned figures must not contain fabricated results or success-shaped mock curves. ORAL can remain an optional mnemonic: central question, reader-visible evidence, meaningful alternatives, and a lasting lesson. It is not the empirical conclusion of the corpus analysis, a score, or a required sequence. Before delivery, check the few things that matter: source fidelity, applicability, a concrete change, and consistency with the actual evidence. Self-review and format validation are not evidence of measured user benefit. Keep the response concise. Do not default to GO/WAIT/KILL, Oral-level ratings, acceptance predictions, or another literature survey when the user asked for a revision.