Draft manuscript sections, format citations, and get AI peer review. Natural language invocation -- describe what you need and the skill routes to drafting, citation formatting, or peer review. Reads research profile, data analysis outputs, and literature search results for context-aware writing. After drafting, offers to generate supporting figures via viz skills. Triggers on: "draft introduction", "write methods", "format citations", "APA style", "bibliography", "peer review", "review my draft", "write abstract". Does NOT trigger for: literature search (use sci-literature-research), data analysis (use sci-data-analysis), image generation (use viz-nano-banana), repurpose/blog/lay summary (use sci-communication).
Transfer the visual style of a reference paper figure onto the user's own data. Takes a screenshot of a figure from a paper plus tabular data and produces an editable matplotlib script in that figure's register, gated by a mechanical quality floor and a Drawer/Reviewer refinement loop. Output is a self-contained .py + PNG + Type-42 PDF. Triggers on: "match this figure style", "figure in the style of", "make my plot look like this paper", "publication figure from reference", "style transfer figure", "reproduce this figure with my data", "NeurIPS-quality version of this chart". Does NOT trigger for: plain data plots with no reference image (use sci-data-analysis), AI-generated illustrations (use viz-nano-banana), flowcharts/architecture/sequence diagrams (use viz-diagram-code), hand-drawn sketches (use viz-excalidraw-diagram).
Publish and edit markdown blog posts on Substack as drafts — converts markdown to Substack's ProseMirror schema, uploads local images to Substack's CDN, pre-renders mermaid diagrams to PNG, and creates or updates drafts via Substack's private API. Draft-only by design — publish is always a human click. Triggers on: "push to substack", "substack draft", "publish to substack", "create substack draft", "send to substack", "substack this post", "update substack", "edit substack draft", "list substack drafts". Does NOT trigger for: reading existing Substack posts, subscriber management, newsletter scheduling. Does NOT publish — only creates/updates drafts you review and publish by hand.
Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
End-of-session checklist that reviews deliverables, collects feedback, fixes skills, updates learnings, and commits work. Use when the user says "wrap up", "close session", "end session", "wrap things up", "we're done", "that's it for today", "session done", or invokes /wrap-up. Run at the end of any working session or after completing a major deliverable. Does NOT trigger for content writing, voice extraction, positioning, or audience research.
Schedule recurring tasks that run automatically via a system-level watchdog, even when Claude Code is closed. Job definitions persist in cron/jobs/ as the registry. The watchdog (launchd on Mac, Task Scheduler on Windows) checks every hour and runs due jobs headlessly via claude -p. Triggers on: "schedule a job", "cron job", "run this every morning", "automate daily", "recurring task", "set up a scheduled job", "check scheduled jobs", "list scheduled jobs", "install watchdog", "uninstall watchdog". Does NOT trigger for one-off tasks or in-session reminders.
Parallel-tempering simulated annealing for rugged multi-modal landscapes where local SA stalls. N replicas at a geometric temperature ladder with periodic adjacent-pair exchanges (Metropolis-Hastings). Contribution-weighted coordinate sampling + O(n) incremental delta evaluators. Triggers: "parallel tempering", "simulated annealing", "PT-SA", "replica exchange", "temperature ladder", "MCMC optimization". Not for: gradient descent, convex optimization, linear programming.
Discrete float64 ULP coordinate-descent polisher for high-precision optimization. Per-coordinate +/-1/2/4 ulp trials priority-ordered by contribution-weighted badness. Use when gradient methods stall at ~1e-12 but true residual is ~1e-13 (spherical codes, kissing configs, lattice packings). Triggers: "ulp polish", "polish to zero", "precision polish", "bridge precision floor", "float64 coordinate descent", "break through 1e-13".