superpower-writing

Use after completing a plan or when .writing/ has accumulated stale data and needs a clean reset for the next task.

Generate and verify BibTeX. Adds a paper by DOI or arXiv id through the Zotero library (it fetches the metadata and open-access PDF), then emits the BibTeX. Verifies a citation's authors, title, venue, and year against the real record to catch hallucinated or misattributed references, and audits a .bib for missing fields, missing DOI or arXiv id, and duplicates. Use when managing citations, generating BibTeX, verifying references, or cleaning up a .bib. 中文触发：管理引用、生成 BibTeX、核验参考文献、整理 bib。

Evidence-reliability check. Walks every % claim tag in LaTeX manuscript/*.tex, confirms \cite{} citekeys resolve against .writing/refs.bib (exported from Zotero), runs semantic match against network-fetched abstracts to catch hallucinated or mismatched citations, optionally checks numeric/table consistency, flags any [NEEDS-EVIDENCE] or draft-only markers. Run on demand or when the skeleton is ready for the human author.

Builds and maintains the project's shared vocabulary (ubiquitous language) in a repo-root CONTEXT.md that persists across papers. Use when pinning down terminology, when the same concept keeps drifting across names in conversation or notes, when a term turns ambiguous, or when idea, outlining, or drafting work names a new system, method, or concept. Pairs with the terminology ledger, which governs surface forms inside one manuscript. 中文触发：定术语、统一叫法、术语表。

Use when proving security of encryption schemes, MACs, signatures, PRFs, hash constructions, or other cryptographic primitives, constructing reduction proofs from standard assumptions (DDH, CDH, RSA, LWE), or writing formal game-based security proofs and advantage bounds in academic papers. Covers game sequences and hopping, negligible-advantage arguments, tightness accounting, and turning an informal security claim into a displayed definition, theorem, and proof a reviewer can verify.

Turns a chosen research area into publishable directions through three internal phases. It generates candidates with named lenses and FINER scoring, adjudicates novelty per technical claim against prior work, then evaluates against a top-venue bar with a fatal-flaw short-circuit. All verdicts stay advisory. Use when a research direction is undecided, when choosing a topic, when evaluating an idea, or when checking novelty (查新).

Find and review academic literature. Use when the user wants to search papers, survey the state of a research direction, look up a paper or dataset, find baselines or related work, enrich a related-work section or comparison table, screen weak citations, or verify that cited works exist. Routes through Zotero semantic search for in-library papers, parallel-cli web search, a deep-research fallback, and direct arXiv, DBLP, Semantic Scholar, and CrossRef access. Chinese triggers: 搜文献, 综述某方向, 查论文或数据集, 找 baseline, 补充对比表, 文献核验.

Use when starting a new paper or when `.writing/outline.md` is missing or needs a major restructure. Converts a research idea into an IMRAD outline plus a per-section claim list, sharpened through an interleaved interview. Produces `.writing/outline.md` (structure + key claims), `.writing/claims/section_*.md` (claim stubs), and a filled `.writing/metadata.yaml`. No manuscript prose is written here, this is the spec phase before drafting.

Read a long PDF (a paper under review, a cited reference) when the answer draws on more than one place in it and embedding the whole file is too big or one-shot. Ships a local no-LLM script for outline, per-page text, and keyword or regex search, plus a subagent fan-out recipe for whole-document extraction. Use when summarizing or comparing sections, locating where a topic is discussed, or listing every dataset, baseline, or citation across a paper including appendices.

Pre-submission visual lint for a compiled PDF: detects text overflowing the page margins, overlapping text or figure blocks, low-resolution embedded images, inconsistent body fonts, and blank pages. Use when the user asks to check the compiled PDF's layout, hunt overfull or margin-overflow problems, verify figure resolution before submission, or says 检查排版 / PDF 溢出 / 图片分辨率够不够 / 提交前过一遍 PDF. Detect-only: it reports page-anchored issues and never edits the source.

Use whenever a writing task will span more than 5 tool calls, multiple sessions, or risks hitting context-window limits, writing a paper, a long literature thread, a multi-section revision, anything where losing mid-task state would hurt. Creates a persistent `.writing/` directory that acts as working memory on disk, so the outline, claims, findings, and progress survive context resets. Every other superpower-writing skill assumes this directory exists; initialize it first. Trigger even when the user does not explicitly ask for "planning".

Polish prose in three passes: strip AI-writing tells, apply Strunk's rules for clarity, then hold every research claim to its evidence. Two modes: edit in place, or diff-first where the original is untouched until the user approves each change. Use when the user asks to 润色 / 改写 / 精修 / 打磨 a draft, or says "polish", "refine", "tighten", or "improve". Preserves meaning, numbers, citations, and [NEEDS-EVIDENCE] markers; never smooths an unsupported claim into a supported-sounding one.

Builds a grounded, venue-compliant response to reviewers for a CS/systems/ML paper. Atomizes every reviewer comment into an issue board with a fixed action vocabulary, enforces three finalize-blocking gates (provenance, commitment, coverage), drafts in strict R-A-C format with page cross-references, and wires every claimed change to a real manuscript location. Use when the user says "rebuttal", "reply to reviewers", "respond to reviews", "OpenReview response", or "ICML/NeurIPS rebuttal".

Bumps the plugin version, tags it, and publishes a GitHub Release with a changelog. Walks a pre-release gate that runs the smoke test, the skill linter, and the eval-harness fixture self-test before any tag is cut. Use when publishing a new version of superpower-writing or preparing a release commit.

Generate raster figure PNGs (architecture diagrams, pipelines, concept art, graphical abstracts) through the local Codex CLI's built-in image tool (gpt-image-2), under a six-section prompt contract with three-attempt selection and a targeted edit loop. Use when the user wants an image-model figure, says 生图 / 用 AI 出一张架构图 / 先出几个方向看看, prefers a polished PNG, or wants layout drafts to pick from before a TikZ rendition. Data plots go to scientific-visualization; routine vector diagrams default to tikz-figures.

Publication-ready data plots for CS papers (systems, ML, networking, security). Use when creating figures with multi-panel layouts, error bars, colorblind-safe palettes, log-scale axes, latency CDFs, throughput curves, training curves, ablation bars, or speedup comparisons for IEEE / ACM / USENIX / NeurIPS / ICML / ICLR submissions. Orchestrates matplotlib, seaborn, and plotly with venue-specific styling and PDF output that drops cleanly into LaTeX. For one-off exploration use seaborn or plotly directly; for system architecture diagrams use tikz-figures or scientific-schematics instead.

Use when proving security of MPC protocols, zero-knowledge proofs, oblivious transfer, commitment schemes, garbled circuits, or any protocol whose security is argued by comparing a real execution against an ideal functionality, in standalone simulation-based or UC frameworks. Covers simulator construction, hybrid arguments, corruption models, composition theorems, and writing the proof section of a cryptographic paper so the ideal-versus-real argument is complete and checkable.

Generate publication-quality LaTeX/TikZ figures for paper manuscripts: architecture diagrams, protocol/data-flow pipelines, sequence diagrams, geometry/math schematics. Compile-verified (xelatex + overlap checker + visual review loop) with a mandatory two-candidate preview for the user to pick from. Use when the user asks for: 论文配图、画架构图、画流程图、TikZ 图、复刻论文图、tikz/latex diagram、vector figure.

User-invoked repair command for a message that did not land. Use when the user types /wait-what after failing to follow the assistant's last message. The assistant re-pitches what it just said, adds the missing context, writes in the conversation's controlled plain register (ASD-STE100 for English, controlled Chinese otherwise), and swaps invented shorthand for the canonical terms in the repo-root CONTEXT.md.