sci-brain

Use when running the autoresearch loop for a project whose survey and validator gates have passed — executes batches of attempts (one git worktree + LOG.md each, scored only by the validator under a hard time limit), then reflects, reports, and re-plans; continues autonomously while authorized_rounds remain, otherwise stops for user review. Stage 4 of the autoresearch pipeline.

Use when setting the goal and building the validator for an autoresearch topic — defines the publishable bar with the user, splits dev instances from a sealed gitignored holdout, builds a Docker-canonical strict environment (recorded fallback allowed), implements the validate CLI with rich JSON errors, and passes the strictness self-test against negative controls. Owns the validator gate. Stage 3 of the autoresearch pipeline.

Use when building the evidence base for a chosen autoresearch topic — maps the insight areas needed to propose new ideas, downloads references to cover them via download-ref, distills each area into research/INSIGHTS.md for user selection, builds a structured domain database and pinned reference implementations, and catalogs all algorithms/software in research/CATALOG.md. Owns the survey gate. Stage 2 of the autoresearch pipeline.

Use when starting or resuming an autoresearch project — reads research/STATE.md to find the current stage of the pipeline (topics → db → validator → run), verifies the previous stage's gate artifacts exist on disk, and routes to the matching stage skill. Triggers on "start autoresearch", "resume autoresearch", "autoresearch status", "what stage is my research project at".

Use when choosing research topics within a domain that suit autoresearch and defining their score metrics — brainstorms candidates, scores each on machine-checkable success, cheap per-attempt evaluation, iteration headroom, and publishability, lets the user pick, then derives primary/guard metrics (with computation cost and gaming risks) for each chosen topic and writes everything to topics.md. Stage 1 of the autoresearch pipeline; invoked via the autoresearch dispatcher or directly with a domain.

Use when building a slide deck in Typst + Touying for a scientific talk, lecture, or project briefing — ships a browsable zoo of color themes, layout templates, and gadgets (callouts, theory boxes, CeTZ diagrams, pin annotations) that compose into projector-safe slides. Distinct from slide-writing (HTML, QudeLeap brand) and paper-writer (full manuscripts).

Use when adding one or many new references (arXiv ID or DOI) to a sci-brain knowledge base — `<project>/.knowledge/` by default, or `advisors/<slug>/.knowledge/` when invoked from an advisor flow. Fetches metadata via Semantic Scholar, downloads PDFs (with SciHub fallback), renders to markdown, regenerates `INDEX.md`, and appends to the KB's `references.bib`. Handles both single refs and bulk-from-bib batches.

Use when brainstorming research ideas — a research collaborator that understands your background, helps find interesting problems together, and shares relevant resources along the way

Use when onboarding a contributor as an advisor — guides them through providing their background and conversation history, runs conversation-dump and soul-extraction, then synthesizes a named advisor profile for the ideas skill's advisor library

Use when reviewing, critiquing, commenting on, or fact-checking an existing scientific manuscript — "comment on my paper", "review my draft", "check my references", "verify my claims", or a referee-style read of a .tex/.typ/.md paper. Not for drafting from scratch (use paper-writer) or for technology/field assessments (use survey-writer).

Use when drafting or revising a scientific manuscript with real results, including journal articles, letters, and arXiv preprints

Use when indexing a paper collection (your own or another researcher's) into a knowledge base — supports Zotero library, a PDF folder, or a Google Scholar profile

Use when surveying a research topic — launches parallel exploration strategies via web search, lets user pick interesting directions, then builds a focused knowledge base with BibTeX

Use when writing up a literature survey, technology review, field assessment, SOTA report, or pros/cons report from a populated knowledge base — the write-up stage of the survey pipeline

Use when attacking a single hard goal that resists a direct solution — runs an autonomous CDCL/DPLL-style search (decide → propagate → learn → backjump, with a meta-pivot when stuck), taking notes after every trial. Domain-agnostic (math, theory, design, debugging, strategy); optionally pulls facts from a project knowledge base. Not for open-ended research ideation (use brainstorm-ideas).

Use when reviewing the visual design/quality of a figure, plot, or diagram — "is this figure well-designed", "review my plot", "critique this figure's layout/colors", "does this figure have good taste". Judges alignment, proximity, color, typography, clarity, and scientific-plot conventions against an 18-rule rubric, renders the figure to look at it, reads its source when available, then prints a scorecard. Report-only (never edits) and terminal-first (saves a file only on request). Not for whether a figure is cited/discussed in a manuscript (use paper-reviewer) or for drafting figures (paper-writer's Figure Rulebook).

Use when writing an ideas report after brainstorming or when turning a chosen research direction into a proposal-style research plan

Use when analyzing conversation patterns — extracts dialog from Claude Code or Codex CLI history, classifies each user message across 6 academic dimensions (Bloom's cognitive level, Graesser question depth, Paul & Elder reasoning probe, Walton presupposition quality, Long & Sato discourse function, Graesser generation mechanism), and outputs tagged dialog reports

Import .md dialog files (Claude.ai exports, custom markdown conversations) to create or update advisor profiles. Use when the user wants to import a conversation file, update an advisor with new dialog data, or add .md chat history to the advisor library. Invoked with /import-dialog.

Use when the user asks to tag, release, or bump the version of the sci-brain plugin (e.g. "tag a version", "cut a release", "bump to 0.3.0", "release v0.x"). Handles the two-file version sync, commit, annotated tag, push, and GitHub release.

Use when extracting conversation patterns and logic jumps from tagged dialog reports — reads /conversation-dump output, clusters trigger→reaction pairs into thinking-pattern.md, detects logic jumps with user confirmation for master-thinking.md