# Academic Title Craft > Generate, compare, audit, and revise titles and method names for ML, AI, and adjacent research papers. Use when the user asks for a paper title, method name, acronym, renaming, title shortlist, or alignment between a title and the paper's actual claim. Do not use for product, company, or campaign names. - Skill: `shichengf/academic-title-craft` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add shichengf/academic-title-craft` - Raw SKILL.md: https://api.skillmd.com/api/skills/shichengf/academic-title-craft/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: shichengf (https://skillmd.com/u/shichengf) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/shichengf/academic-title-craft --- # Academic Title Craft Create a title that identifies the paper's scientific contribution without claiming more than its evidence supports. Memorability is useful only after the title is accurate and natural. ## Establish the title contract Inspect the manuscript or supplied summary before proposing names. Determine: 1. the paper's central question and one sentence answer. 2. the contribution type, such as a method, empirical finding, theory, benchmark, resource, or position. 3. the strongest claim supported by completed evidence. 4. the object being studied and the scope that limits the claim. 5. the role of any named architecture or representation in the contribution. Do not infer that a frozen component is continually learned. Do not turn a state estimator into a dynamics model or a query interface into planning. If the evidence is unfinished, title the stable problem and mechanism rather than an expected result. ## Choose one foreground A title should foreground one of these elements: - the scientific problem or object. - the main supported finding. - the mechanism that makes the result possible. - a useful method name followed by a plain description. Do not compress the abstract or contribution paragraph into the title. Details that explain a secondary learning loop, loss term, benchmark, or implementation boundary usually belong in the abstract. ## Generate title families Draft candidates from more than one family before ranking them: 1. **Direct description:** mechanism or object followed by the task and scope. 2. **Claim statement:** the strongest supported finding in ordinary language. 3. **Method name plus description:** a distinctive name, then a clause that remains informative without the name. 4. **Grounded contrast:** a short contrast that expresses the actual technical tension demonstrated by the paper. 5. **Question:** use only when the paper is organized around resolving that question and a statement title would conceal the answer. Generate acronyms after a natural title exists. Reject an expansion that uses awkward grammar, irrelevant words, or misleading terminology merely to obtain desired letters. The descriptive part must still work when the method name is removed. ## Audit every candidate Reject a title that fails any of these gates: - **Scientific fidelity:** every capability in the title is implemented and evaluated or clearly presented as a formulation. - **Scope calibration:** the title names the actual regime when a broader term would imply unsupported generality. - **Natural syntax:** a reader can say the title aloud without repairing its grammar. Avoid ownership chains, serial enumerations, and noun stacks. - **Role alignment:** the method name describes the component that actually changes or learns. - **Searchability:** the title contains the terms that the intended research community would use to find the work. - **Distinctiveness:** an exact title and method name search reveals no obvious collision with a closely related paper, benchmark, workshop, project, or package. Treat 8 to 14 content words as an empirical prior rather than a limit. Shorter is not better when it removes the object or scope. Longer is justified only when each added phrase prevents a real misunderstanding. Avoid unearned words such as `novel`, `general`, `unified`, or `foundation`. Avoid a broad field label that does not distinguish the paper. Do not repeat the same concept in a method name and its subtitle. ## Check collisions when a name may be public Search the exact quoted title and method name. Also search close spelling and acronym variants in arXiv, OpenReview, official proceedings, and code hosting sites when available. Prefer primary records. State that the check is preliminary because absence from search results does not establish ownership. ## Rank and present Score candidates with scientific fidelity as the first priority. Then compare scope precision, readability, contribution focus, searchability, and memorability. A clever title cannot compensate for a scientific mismatch. Present one recommendation and at most two meaningfully different alternatives unless the user requests a longer brainstorm. For each alternative, explain what it foregrounds and its main risk. Say directly when the existing title is better than the new candidates. Do not edit manuscript files until the user chooses a title or explicitly asks for the rename. After an authorized rename, update the title, method macro, abstract, introduction, captions, conclusion, supplementary material, and project memory wherever the old identity remains. ## Evidence and venue patterns Read [references/title-patterns.md](references/title-patterns.md) when the user asks for evidence, venue comparison, acronym design, collision analysis, or a high stakes final choice. It records the recent top conference corpus, title families, award examples, and the limits of the observed patterns.