horef-claude-plugins

Skills and tools that improve themselves from use: observations, evidence-gated reflection passes, evals, versions and a session-start digest

1 skill · hooks

Lecture or tutorial slides in DEVONthink into complete, self-contained, Anki-ready teaching notes with figures (uses scientific-figures)

1 skill

Design GUIs people actually use: tool and mod panels, settings pages, inspection viewers, dashboards and interactive explainers, at the right density (everything needed on screen or one click away, rarer flows behind more steps, more detail on request). Layered disclosure rule, settings and explainer patterns, surface limits, a build-lint-shoot-look loop with screenshots at several widths in light and dark, and a profile for audience and taste. Use when building or reworking any GUI, panel, page, viewer or interactive demo, or when one feels too complex or too thin.

1 skill

Delegate self-contained steps to helper agents on other subscriptions (Kiro, Antigravity/agy, GitHub Copilot, Codex, another Claude Code account) through Herdr's herdr-core (hc): pick a helper by tier, launch it with a brief, act as its approval judge, collect its report, review it, clean up. Use when the person asks to offload or delegate work, when Claude's limits are tight, or when a router plan marks a step for herdr. Needs hc on PATH.

1 skill

Literature research and claim verification for any research question: frame the question and the decision it serves, search several sources in parallel (PubMed, Semantic Scholar, arXiv, OpenAlex, the web, DEVONthink when on the owner's Mac) with a reproducible search log and citation chasing, record every finding in an evidence ledger (value, units, species, preparation, source, page, verbatim quote, evidence type) before any prose, have fresh checkers re-read each cited passage, then synthesise the answer with conflicts and costed gaps. Use for parameter lookups, mechanism or state-of-the-field reviews, novelty or prior-art checks, verifying a claim or citation, and finding methods; not for writing the final paper (use research-writing).

1 skill

Run many local simulation or analysis jobs in parallel under a private-memory budget: measure each job's whole process tree (macOS footprint, Linux PSS; never RSS), admit jobs by estimates taken from measured peaks, stop and requeue the newest job above a kill line, chain dependent jobs, survive restarts and the launching agent's exit, and change budget, priorities or estimates while it runs. Use when a batch of local runs (sweeps, calibrations, simulations) must fit a machine's memory, when parallel runs need a memory estimate, or when a previous queue over- or under-used memory.

1 skill

Portable model routing as a skill for agents without the model-router mod (Copilot, Kiro, Codex, Cursor, claude.ai): plan steps by tier, get approval, run each on the cheapest model that does it well

1 skill

Pre-read a scientific paper in DEVONthink and mark it up in Sergiy's colour scheme (whole logical units, comments, skim-safe markers) plus a linked reading-map note. Works with or without a project.

1 skill

Set up, retrofit, upgrade or audit a repo so agents and humans can maintain and share it: AGENTS.md, status, decisions, generated map and checks

1 skill

Run a research question through the whole loop with checkpoints: frame it and the decision it serves, review the literature (literature-research), turn findings into competing hypotheses and a costed analysis plan, run the approved analyses (bounded local runs; large compute only on a yes), record results in the project, and write them up (research-writing). Stops for the person at three gates: after framing, after the literature (which hypotheses to test), and before any compute. Starts from a question in a project, a new idea (novelty check first, then a scaffolded project), reviewer questions, or a paper to build on. Use when asked to research, investigate or test an idea end to end, to plan analyses for open questions, or to take a question from literature to results; for a literature question alone use literature-research, for writing alone research-writing.

1 skill

Turn a body of research (results, run logs, figures, decisions) into a coherent write-up or paper draft in LaTeX: a one-paragraph message and its reader, a claims ledger tying each claim to its evidence and figure, a skeleton in argument order (not lab-notebook order), figures that each carry one claim, text written around them, then a loop of fresh reader agents whose questions are answered in the text, plus mechanical checks and a look at the rendered pages. Works for a full write-up after many results or incrementally as results arrive. Use when asked to write up results, draft a paper, report or thesis chapter, revise one that reads like a lab report, or add new results to an existing write-up.

1 skill

Scientific figures for notes, papers and slides (circuits, anatomy, data plots, phase portraits, 3D surfaces) with a lint-render-look loop; a personal profile sets audience and output targets

1 skill

Routes each step of a task to a model tier (light, standard, deep), with plan approval, a status line and a plan pane

A compact status line above the prompt: model and effort, context fill, 5h and weekly rate limits with reset and pace, Kiro, Copilot and Codex quotas, your other Claude accounts' limits and Claude extra usage, cost, last turn, compactions and files edited; one line, or two when narrow