p135246/computational-research

Create and manage work items in the top-level Work/ folder — each a multi-session effort with a Spec (what to build), a Tasks checklist (one task ≈ one session), and a Progress log. This is the project's execution state, separate from the Wiki knowledge base. Use for: "new work item", "start a work item", "spec out X", "plan for X", "break X into tasks", "track this across sessions", "add a task", "update the spec", or the /backlog-add command. Creates Work/Backlog/a named item.md, bootstrapping the folder if missing. A work item's status is its folder (Backlog/Ready/Active/UnderReview/Done/Dropped), not a field. Specs follow the document-revise protocol; shaping one together with the user over a long sitting is the backlog-refine skill. Do NOT trigger on casual uses of the word "work".

Execute approved research backlog tasks autonomously in isolated worktrees, using native subagents when available and reporting a digest for human review.

Create a standalone local research backlog board with human edits exported for review and reconciliation.

Go through the whole Work/ backlog with the user: read every item in Backlog/, Ready/ and UnderReview/ and every work/a named item branch, and report what waits on the user (a review, a halted run's question, a Waiting-on line, queued board changes), what is stale (no Progress for a month, or a premise the code no longer has), which items overlap, and what to refine, merge, split or drop next. Changes nothing until the user says so; then each move is a git mv and an index update. Never moves an item to Ready and never syncs the board. Use for: "go through the backlog", "backlog info", "what's in the backlog", "what waits on me", "clean up the backlog", "what next", or the /backlog-info command. Not for shaping one item (backlog-refine) or reviewing a finished one (backlog-review).

Shape a Work/Backlog item together with the user over a long sitting until it is ready to be worked — the user writes the human parts (Summary, Motivation, Acceptance criteria, any detail they care about) in their editor, while the LLM researches the code and wiki, writes the Technical details, keeps a list of open questions, and proposes the tasks and their routing. Ends with a readiness check and, on the user's word, the move to Work/Ready/ where /backlog-run-scheduled may pick it up. Use for: "refine X", "let's work on the backlog", "shape this item", "backlog refinement", "get X ready", or the /backlog-refine command. Not for executing tasks (backlog-run) or filing a quick new item (backlog-add).

Review a finished Work/ item with the user: find it in Work/UnderReview/ or on its unmerged work/a named item branch, show the run digest, walk the Acceptance criteria one by one with the test instructions of the tasks that serve each, run the checks that can be run, and end on the user's verdict — accept (merge the branch, item to Done/) or send back (merge the branch, a new task from the user's words, item to Active/). Nothing is fixed during the review. Use for: "review X", "review the item", "check the finished item", "accept X", "send X back", "merge the item branch", or the /backlog-review command. Not for running tasks (backlog-run) or going through the whole backlog (backlog-info).

Run one disciplined work session against a Work/ item: pick the next incomplete task, study the Spec, implement exactly ONE task, append a Progress report, mark the task done, sync durable knowledge to the Wiki, commit, and STOP. Built to be run in a FRESH session each time to avoid context accumulation. Use for: "next session", "next task", "work the next task", "continue the work item", "resume a named item", or the /backlog-run command. It does exactly one task then stops — it never chains tasks.

Run an approved backlog item unattended through bounded Codex CLI sessions in an isolated worktree.

Review agreement between a current human-owned manuscript or paclet design, its notes companion, and relevant proofs, code, examples or checks. Use when the user asks whether implementation follows the accepted design, which proposals remain unintegrated, or where claims and evidence diverge. Report focused discrepancies with locations and evidence; do not rewrite the human baseline, accept proposals or advance its iteration during review.

Maintain one concise research companion to a human-owned manuscript or paclet design. Use when the user asks for itemized proposals, mathematical statements, API suggestions, corrections, questions, citations, or updates to a manuscript's or design notebook's _notes companion. Preserve the human baseline and human corrections; put authorized lengthy investigations in linked topical artifacts rather than writing a replacement paper or design.

Apply human feedback and direct edits to a document within the requested scope, preserving unrequested content. Use for "revise this", notes left in a notebook, and /document-revise. Keep human manuscripts and _notes companions in place unless a new iteration is explicitly requested. A requested versioned revision round preserves the prior document and records notes, hand edits and outcomes in its provenance file.

Locate the current human-owned paper or paclet design and its concise notes companion, or establish that pair during requested setup. Use when the user asks to set up human-led paper/design collaboration, select the current manuscript or design iteration, or identify which design implementation should follow. Do not use ordinary research or build requests to invent a new human iteration.

Create Wolfram notebooks from Markdown, edit saved .nb files through notebook-aware tools, or export requested Markdown without changing the notebook. Use for "create notebook", "edit notebook", "update notebook", "put this in a notebook", and notebook/Markdown conversion. Preserve direct human edits, metadata and established paper or paclet design paths. Other skills use the creation pipeline when generating new notebook cells.

Build a Wolfram paclet archive and install it locally, consulting the current human-owned design notebook and concise _notes companion when present. Support the design-document workflow during paclet development. Use when the user says "build paclet", "build the paclet", "install paclet locally", "make paclet", "create paclet archive", or "test the paclet install".

Generate real Wolfram documentation for a paclet — one symbol reference page per exported function, written into the paclet's Documentation tree through the official MCP doc tools so it resolves in-product via ?Symbol, F1, and the Documentation Center. Use when the user says "document the paclet", "paclet docs", "symbol pages", "reference pages", "document these functions", "write docs for X", or the /paclet-docs command. Guide pages are deliberately not generated — they are a human deliverable.

Build a Wolfram paclet, install locally, and publish to Wolfram Cloud with a public install URL. Use when the user says "publish paclet", "upload paclet", "deploy paclet", "release paclet", "push paclet to cloud", "share the paclet", or "make paclet available".

Produce a BibTeX/biblatex entry from an arXiv ID or DOI. Optionally appends to Research/references.bib if it exists. Use when the user says "cite this", "bibtex for arXiv:...", "DOI 10.xxx/...", "add citation", or when adding a reference during paper writing.

Scaffold a human-owned LaTeX or Typst paper and its concise _notes research companion, preserving an existing project's folder and naming conventions. Read the current human manuscript before proposing work; edit it only on an explicit request. Default to itemized proposals, scoped claims, questions, citations and artifact links, not a full model-authored replacement paper. Use for paper setup, LaTeX/Typst scaffolding, and collaboration on a human paper.

Turn a conversation into a concise mathematical note or requested code extract, preserving an existing paper's format and ownership. Use for "research note", "write this up as a note", "make an artifact of this", "code extract", or /paper-create-note. Distinguish proofs, assumptions, computations and open questions. Produce only the requested deliverable; the complete LaTeX/notebook/code bundle is available when explicitly requested. Keep the note local; nothing is uploaded to Wolfram Cloud.

Write an explicitly requested mathematics paper or article as a Wolfram notebook, with complete proofs, cited sources, numbered statements and executable examples. Use for an explicit request for a paper as a notebook or /paper-create-notebook; ordinary exploration, conjecture notebooks and demonstrations use notebook-create. Generate from Markdown with MathNotebook environments and preserve detected notebook edits. Human-owned papers follow their ownership policy. Publish to Wolfram Cloud and add public README links only when the user explicitly authorizes publication of this notebook.

Optionally maintain a running scientific journal in LaTeX or Typst — a concise, structured, append-only stream of dated def/thm/rem/claim entries recording the math and physics content and main claims established during research, with every resource cited. Off by default; gated by a per-project toggle in AGENTS.md. Scaffold it, append dated entries automatically when on, and list the index. Use when the user says "keep a journal", "scientific journal", "journal this", "turn the journal on/off", "journal status", or the /paper-journal command. Distinct from the Wiki (deduplicated encyclopedia): the journal is the typeset, cited record of what was learned and when. It is also where everything a paper cannot carry goes — hedged claims, verification ranges, heuristics, alternate proofs, failed attempts and unresolved [lookup] citations — so nothing is silently dropped. Other skills (paper-create-notebook, paper-create, wiki-update, backlog-run, paper-cite) feed it when on.

Drive a Lean 4 / Mathlib formalization session via the lean-lsp MCP. Opens a file, reads the goal state, searches Mathlib for closing lemmas, verifies proofs, and writes a formalization checklist to Work/. Use when the user says "formalize in Lean", "prove this in Lean", "what's the Mathlib name for X", "check this Lean proof", "lean state", or during math-research projects with a Lean/ subdirectory.

Check local research tools, connected Wolfram MCP capabilities and license headroom without starting a kernel.

Archive superseded project materials when cleanup is requested, preserving human work and annotated reading copies.

Scaffold a new research, math-research, paclet-dev, or paclet project and initialize its research infrastructure. Use when the user explicitly asks to create a project, set up a project folder, scaffold a new paclet or dev repo, or invokes /project-create. A request to investigate a topic or run a one-off computation does not by itself request project setup. For an existing project, use the skill for the requested missing piece rather than scaffolding again.

Summarize an existing research project and its next tasks.

Optionally record the prompts/intent behind generated artifacts — notebooks, paclet functions/code, wiki articles, and work items. Maintains an append-only ledger at Wiki/Prompts.md and embeds a back-pointer in each artifact so provenance travels with the file. Off by default; gated by a per-project toggle in AGENTS.md. Use when the user says "track prompts", "record the prompt", "where did this come from", "turn provenance on/off", "show the prompt ledger", or the /project-provenance command. Other skills (notebook-create, wiki-update, backlog-add, backlog-run) follow this skill's format when the toggle is on. Also owns the always-on document provenance file (stem.provenance.md) that a revision round keeps beside a document: notes, hand edits, requests, rules.

Start, resume, or manage a guided tour of the project. Creates an interactive walkthrough with narrative and runnable code for each topic. Use when the user says "start tour", "give me a tour", "continue tour", "where were we", "tour status", or when presenting the project to someone.

Add a resource (paper, repo, notebook, dataset, tool, web page) to the wiki. Downloads the file locally (gitignored), creates a wiki article with a Recover section for reproducibility, and updates the index. Use when the user says "add this paper", "save this resource", "note this repo", or when encountering a new reference during work. Also use during project-create to register discovered papers and Wolfram Community resources.

Run a health check on the wiki. Finds stale articles, missing coverage, broken backlinks, and opportunities for cross-linking. Use at the start of a new session, or when the user says "check wiki", "wiki health", "audit wiki", "wiki status".

Initialize a plain-markdown knowledge base (Wiki/) in the current repo. Use when the user says "init wiki", "set up knowledge base", "add wiki to this repo", or when starting a new project that needs a wiki. Also triggers when wiki-update or wiki-check is invoked but Wiki/ does not exist yet.

Search external mathematics resources for a topic: MathWorld, nLab, OEIS, DLMF, and Wikipedia math articles. Returns ranked results, optionally hands them to wiki-add-resource. Use when the user says "search math", "find on MathWorld", "look up in nLab", "OEIS sequence", "DLMF formula", "search Wikipedia math", or during math-research project setup when external math sources are needed.

Search Wolfram ecosystem resources for a topic: documentation, Function Repository, Community forum, Stephen Wolfram's writings, and the Wolfram Physics Technical Introduction. Downloads notebooks, creates wiki articles, and logs findings. Use when the user says "search wolfram", "find wolfram resources", "wolfram community search", "what functions exist for X", "check the function repository", or during project setup when the user wants Wolfram-specific resource gathering.

Update the wiki after a substantial change. Creates or updates articles, maintains the index, status, and backlinks. Use after any significant step — code changes, new functionality, completed tasks, discoveries. Triggers on: "update wiki", "log this", or automatically after substantial work. Also invoke proactively after completing any multi-step task.