Agent Automation & Development Skills

Use when drafting prose that DEFENDS or DEVELOPS specific claims using cited evidence from the local library — claims the user has stated explicitly, or claims they want help articulating with grounded support. Covers decomposition into grounded propositions, per-claim invocation of the grounding-against-library kernel, and structured output with prose + auditable per-proposition appendix. NOT for topic exploration ("summarize what the literature says about X", "overview of approaches to Y", "tour what these papers cover") — use the using-local-library-mcp orientation skill for those, which produces well-quoted grounded prose without claim-decomposition. Fires on phrasings like "defend the claim that...", "develop the argument that...", "extend this thesis with evidence from...", "make the case that...", "write a paragraph supporting [proposition] using @ ", "draft a section arguing [position] from @ and @ ".

Use when grounding exactly ONE target assertion against library evidence — retrieves via MCP tools, widens context as needed, synthesizes across all retrieved evidence with per-source tagging (support/contradict/partial/not-found), and reports with per-source citations and quoted evidence. Also use for verification prompts like "verify the claim that...", "is it true that @ claims X", "check whether @ says Y", "ground this assertion...". Callers must decompose broad requests into atomic assertions and invoke this skill per-assertion; the kernel returns control if the target is too broad.

Use when the user asks about an equation, formula, derivation, loss function, algorithm, figure, table, or other visually-formatted technical content from a library document — Marker's PDF extraction generally produces correct LaTeX for math, but it can degrade with subscript/superscript collapse (QWQ i instead of QW^Qi), escape-sequence remnants like \n(1) at equation tails, empty math delimiters, empty markers, broken markdown tables, or showdocument reporting Status: needsreview. Default flow: retrieve the markdown, inspect for those genuine garbling cues; if cues fire, escalate to the PDF via Read at the path from showdocument. Provenance is mandatory either way: cite chunk index for markdown-sourced content, page number for PDF-sourced content. Do not ask the user to retype.

Use when comparing code or implementation against method descriptions in one or more source papers — covers decomposition of code into checkable claims (algorithm shape, hyperparameters, loss/objective terms, normalization, initialization, edge cases), per-claim grounding via the grounding-against-library kernel, math-verification fallback to the PDF via the handling-extraction-quality skill, and an assertion-table output mapping each code claim to its paper locus and a status (match / mismatch / partial / not-found). Multi-source. Also fires on phrasings like "compare this code/file against @ ", "does this implementation match the paper", "audit my implementation of ", "verify that this code follows @ ", "check whether the hyperparameters match", "does this loss function match what's in @ ".

Use when citekeys (@Name2023, bare Name2023), paper/book/chapter references, phrasing like "what does X argue", "from my library", "cite this", or any request that could plausibly be grounded in the user's local research corpus — reaches for mcplocal-library tools before answering from training, and treats @citekey as a library handle rather than a topic label.

Use when checking whether the local library supports given textual claims — claims with explicit citekeys, claims attributing positions to named authors, or claims about what a source "argues" / "shows" / "finds" / "concludes". Enumerates each citation-bearing claim, invokes the grounding-against-library kernel per claim, and reports a claim table with verbatim quoted support and a status (match / partial / mismatch / not-found). Multi-source. Catches "doesn't contradict" passing as "supports", silent qualifier-stripping, and one-chunk-dispositive errors. Also fires on phrasings like "verify these claims against my library", "fact-check this passage", "are these citations warranted", "does my library actually back these claims", "is @ really saying X", "check whether the cited sources support this".