Use when someone is about to lose access to a machine (shared server, lab box, cloud/VM instance, expiring rental) and needs to preserve all their work before it's gone. Systematically finds everything that exists ONLY on that machine — unpushed commits, stashes, untracked/uncommitted files, dirty submodules, non-git dirs, loose files, and large gitignored artifacts — and backs it up to GitHub / HuggingFace / an off-machine archive, non-destructively. Triggers on "losing access to this machine", "back up everything before I lose the server", "migrating off this box", "decommission", "my instance expires".
Use when preparing a paper's camera-ready or arXiv release — de-anonymizing an accepted submission, packaging an arXiv tarball, checking dual-submission safety against a concurrent venue (e.g. a paper accepted at a workshop that's also under review at NeurIPS/ICML), or flagging reviewer-requested post-review additions. Encodes a byte-identity discipline (the public version changes ONLY author/venue metadata, never content), a numbers-from-scripts integrity gate, and the arXiv mechanics that actually bite.
Spin up one or more fresh-context subagents to critically evaluate a proposed theory, hypothesis, interpretation, or experimental design before committing to it. Use when the user says "have an agent critique this" / "get a second opinion on" / "is this design sound" / invokes /critique, or proactively before running any experiment or committing to a non-trivial theoretical claim.
Use when launching a multi-hour neural-network training, fine-tune, or other long GPU job autonomously from Claude Code and you need to catch failures (NaN, stuck-at-chance, dead process, throughput collapse, OOM) early instead of waking up to a wasted GPU window.
Use when the user asks to check, audit, or improve a website or web project for accessibility (a11y), WCAG compliance, screen reader support, keyboard navigation, color contrast, or alt text. Triggers a plan-mode investigation against the TeachAccess design and code checklists, then implements approved fixes.
Consolidate scattered research notes, logs, experiment outputs, and submodule docs into a single living research paper. Use when the user wants to pull together multiple source documents into one structured paper.
Use when the user asks to review, find, summarize, or check Claude Code chat transcripts from a past date or time range ("review my chats from May 1st", "what was I working on yesterday", "any unfinished sessions this week"). Reads transcripts under `~/.claude/projects/`, handles local-time vs UTC correctly so late-evening sessions don't get dropped, and flags chats whose last assistant turn looks like an unanswered question.
Translate math (formulas, estimators, algorithms) into code so the implementation faithfully matches what the source actually specifies. Use when writing code from a formula, reviewing an LLM-generated implementation of a formula, debugging a numerical mismatch with a paper, designing a new metric/estimator, or refactoring an existing math-heavy computation. Especially load-bearing whenever aggregation operators (sums, means, expectations, products, geometric means) appear over indices that can be reordered, or whenever the same English label can refer to multiple non-equivalent estimators (e.g. ratio-of-means vs mean-of-ratios, micro-average vs macro-average, sample-weighted vs unweighted). Prevents the failure mode where a code path silently implements the wrong estimator under the same name as the intended one.