claude-skills

Hierarchical review of scientific documents (papers, grant proposals) without editing them. Layer 1 spawns naive skim and adversarial critics that file findings as GitHub issues; Layer 2 triages each finding simplification-first (cut before patch). Runs coarse to fine — document skeleton, sections, paragraphs — locking user decisions between stages, then a sentence-level style check. Use for any review or critique request on paper or proposal prose. For writing, rewriting, or wording edits, use the scientific-writing skill instead.

Draft and revise multi-page scientific documents (papers, grant proposals) in the user's style. Structures documents for busy readers who skim — skeleton-first drafting loop, claim-carrying headings — and enforces the user's sentence-level voice (terse first-person description over persuasive, blog-like phrasing). Use whenever writing, rewriting, or editing paper or proposal prose, even for a single sentence. For a review or critique request without editing, use the scientific-review skill instead.

Set up an orchestrated lead-agent + subagent loop for any iterative multi-step task. Use this skill whenever the user wants to process a collection of items one at a time (papers, files, tickets, data points), build up accumulated results across iterations, or run a multi-phase pipeline with a controlling agent and worker subagents. Also trigger on "ralph loop", "lead agent loop", "orchestration loop", "subagent delegation", processing a list of things automatically, or iterating over a batch of inputs. Trigger even when the user just describes the task without naming the pattern—if it sounds like "do X for each Y and accumulate results", this skill applies.