marcus-waldman

Audit the mi-spectral sourced derivation (manuscript/derivation.qmd, or any manuscript/*.qmd) for verification and citation-discipline compliance, and emit a flag-only report. READ-ONLY: it diagnoses and flags, it never edits the manuscript or literature. Use this whenever the user asks to audit, check, verify, vet, or review the derivation / the proof / the sourced derivation for citation integrity, broken source pointers, hallucinated or unsupported citations, unreproducible numbers, dangling cross-references, missing Monte Carlo asserts, or over-claiming — and when they say "/derivation-audit", "audit the derivation", "is every claim sourced", "check the proof's citations", or "does the evidence actually back the claims".

Run one cycle of the human-in-the-loop manuscript review protocol on a target (level1, level2, level3-decomposition, or a section id like S4). Use when the user says "/section-review <target>", "review section X", "let's do the level 1 annotation", "start the review loop", or returns an annotated extract containing

Project-manager dashboard for the mi-spectral manuscript. Use when the user asks "/status", "where are we", "what's next", "what am I blocked on", "what's waiting on me", or at the start of a manuscript work session. Derives the picture from manuscript/plan/status.json, decisions.md, the level files, and git — never from session memory — and reports drift if they disagree.

Sync the Better BibTeX export to mi-spectral/literature/<citekey>.md files (citation-discipline bridge). Use when the user asks to "sync the literature", "refresh literature files", "lit-sync", or after exporting a new BibTeX from Zotero. Parses literature/mi-spectral-better-bibTeX.bib, finds each PDF in the MI-SPECTRAL drive folder by author/year, looks up cached Mathpix conversions by SHA256, and writes per-citekey markdown with YAML frontmatter. Idempotent.

Generate a PRISMA-style flow diagram for the NE25 analytic-sample construction (REDCap pull → eligible → analytic sample → weighted), with human-in-the-loop checkpoints to confirm exclusion-stage layout and methods-section wording. Output is a Mermaid `flowchart TD` block, either pasted into chat or written to a target `.qmd` file. Invoke when documenting the NE25 sample for a methods section, the WEIGHT_CONSTRUCTION narrative, an audit trail, or a collaborator handoff.

Regenerate `docs/index.html` (the GitHub Pages onboarding/orientation landing page) from current platform state — DB counts, CLAUDE.md status, recent commits, and known drift items. Invoke when the platform changes materially (new pipeline added, deferred work shifts, major status changes, DB drift items resolved or new ones surfaced) so the public-facing snapshot stays accurate.

Surgically update `HANDOFF.md` (root) — the maintainer-transition synthesis doc — so its volatile fields (in-flight work, drift items, last-audit commit, snapshot date, uncommitted-work breadcrumb) reflect current reality. Invoke during the pre-handoff window whenever material changes land that affect what the incoming maintainer needs to know. Sister skill to `/refresh-onboarding` (which regenerates the public Pages page); this one updates the internal markdown.

Regenerate `docs/database/TABLES.md` from live DuckDB introspection + the hand-written metadata in `docs/database/table_metadata.yaml`. Invoke when a table is added, dropped, renamed, moves between pipelines, or when row counts materially drift. Third member of the docs-refresh skill family (alongside `/refresh-onboarding` and `/refresh-handoff`).

Multi-prong, context-efficient Q&A over the Iconsult knowledge-graph corpus (arsanjani_2026 + gulli_2025 + bratanic_2025). Decomposes a question into facets, fans out ISOLATED retrieval subagents (concept match + graph traversal + passage RAG across all three books via the canonical layer), adversarially verifies each synthesized claim against its cited passage, and returns a tight, cited answer — WITHOUT dumping the bulky intermediate retrieval into the main conversation context. Use when the user has a substantive question to ANSWER from the books/KG. E.g. "what does the corpus say about X", "how does reflection relate to evaluator-optimizer", "/ask-corpus <question>". Do NOT use to score/consult on a user's architecture against the rubric — that is the heavyweight `consult` MCP prompt's job.