Skills Collection

A curated collection of reusable skills for common automation, research, and infrastructure workflows.

Overview

This repository groups a small set of focused skills into one place with consistent naming and structure. Each skill lives in its own directory, includes a SKILL.md entrypoint, and may also bundle scripts/ or references/ depending on the workflow.

Output directory convention

Skills that write generated artifacts into a project's outputs/ directory name each run directory YYYYMMDD_<topic>, with the date taken from date +%Y%m%d on the day the directory is created. Alphabetical order then matches chronological order, so a bare ls outputs/ reads as a timeline and no separate sequence index is needed.

Skill	Directory
research-report	outputs/YYYYMMDD_<topic>/; the same basename is reused for the ~/Dropbox/ResearchReport/<Category>/ archive
adversarial-review	outputs/review/YYYYMMDD_HHMM/, with a time suffix because one draft is often reviewed more than once a day
overleaf-section-workflow	outputs/YYYYMMDD_<topic>/ in the research project that owns the plot scripts
concept-explainer	YYYYMMDD_<concept-slug>/ in the working directory, mirrored under the same name to ~/Dropbox/ConceptExplainer/<Topic>/

Directories created before this convention keep their names, and a directory is never re-dated because it was edited later: the prefix records when the work started, which is what makes the ordering stable.

Included skills

Skill	Primary use	Entry point	External setup
academic-jobs	Fetch valid (still-open, deadline-not-passed) academic job postings from Academic Jobs Online (AJO) and the InspireHEP jobs board via the ajo CLI: searches both boards by the same field presets (keywords + position-type + country filters + which sources), fetches and stores postings judged valid from their effective deadline, flags what is new since the last check, and inspects a posting. Postings keyed by (source, id) in a local SQLite store under ~/.local/share/academic-jobs/.	academic-jobs/SKILL.md	uv (the bundled ajo uv project auto-installs requests + beautifulsoup4 on first uv run)
adversarial-review	Stress-test a paper draft or report with a parallel persona swarm (hostile theorist, statistician, editor, citation auditor, figure critic) and produce a ranked fix list	adversarial-review/SKILL.md	None
bibtex-gen	Generate bibtex entries by routing each reference to its most authoritative source — InspireHEP for HEP, Google Scholar (via scholarly) for non-HEP, CrossRef DOI bibtex as the publisher fallback. Auto-classifies HEP via an InspireHEP probe; --hep / --no-hep for overrides. Accepts arXiv IDs, DOIs, titles, or URLs and supports batch input.	bibtex-gen/SKILL.md	None — scholarly is declared in the orchestrator's PEP 723 header and auto-installed by uv run
commit-triage	Classify uncommitted changes into commit / failure-archive / ambiguous buckets and produce clean grouped commits with no co-author attribution	commit-triage/SKILL.md	None
concept-explainer	Explain a specific concept (physics / math / ML / stats / CS) to a named free-form audience with full mathematical rigor and many visualizations — explanation.md + executable matplotlib plot scripts (scienceplots ["science", "nature"], never no-latex) + optional Friendly Whiteboard schematic prompts. Auto-renders PDF via md2pdf-typora; every explanation is archived to ~/Dropbox/ConceptExplainer/<Topic>/YYYYMMDD_<concept-slug>/.	concept-explainer/SKILL.md	matplotlib, scienceplots, system TeX install (for LaTeX rendering); md2pdf-typora for the PDF step; optionally wide-slide-illustrator / codex-image for schematics
dropbox	Upload, download, and share files through the Dropbox API	dropbox/SKILL.md	OAuth credentials (interactive)
hep-rumor-mill	Analyze the HEP-theory postdoc rumor mill (a public Google Sheet) via the prm CLI: pull a year's offer list, resolve each offer-holder's InspireHEP record (plus OpenAlex by ORCID and Semantic Scholar by name for interdisciplinary people), and study what kind of profile lands where. Institute cohort profiles and self-benchmarking against the accepted cohort, with Korean report output. State in a local SQLite store under ~/.local/share/hep-rumor-mill/. Self-reported data, treated as descriptive not predictive.	hep-rumor-mill/SKILL.md	uv (the bundled prm uv project auto-installs requests on first uv run)
handdrawn-schematic	Generate a single friendly hand-drawn whiteboard schematic (wavy marker strokes, numbered panels left-to-right, chunky chalk arrows, hand-written notes) on a pure white background that explains a concept, pipeline, architecture, or algorithm at a glance. Composes the load-bearing style block + a 3-6 panel figure brief and by default renders a PNG via the bundled codex image_generation tool (ChatGPT OAuth); falls back to emitting the copy-paste prompt when codex is unavailable. The reusable single-figure generator extracted from journal-club-review, retuned from cream to white.	handdrawn-schematic/SKILL.md	Logged-in bundled codex for rendering (optional; prompt-only without it)
journal-club-review	Turn an arXiv id/URL, a PDF, or raw text into a journal-club paper presentation (nine sections: TL;DR, The Problem, Key Idea, How It Works, Key Results, Why It Matters, Strengths/Limitations/Open Questions, Discussion Questions, Takeaways) meant to help a reading group understand and discuss a paper, not produce a referee report. LaTeX math, source-language auto-matching, and two optional friendly-whiteboard figures (method + results).	journal-club-review/SKILL.md	uv (extractor deps auto-installed via PEP 723); logged-in bundled codex for figures (optional)
md2pdf-typora	Convert Markdown to PDF that mimics Typora's Whitey-theme export (pandoc + Chrome headless, MathJax SVG, Korean serif fallback)	md2pdf-typora/SKILL.md	pandoc + Chrome/Chromium
morgen	Manage calendars, events, tasks, and tags across Google/Microsoft/iCloud/CalDAV accounts via the Morgen API	morgen/SKILL.md	Morgen API key
overleap	Bidirectional real-time sync between an Overleaf project and a local directory via the overleap Node.js CLI	overleap/SKILL.md	overleap CLI + Overleaf session cookie
overleaf-section-workflow	Disciplined section-by-section workflow for Overleaf physics-paper drafts: Korean intermediate draft → user iteration → Opus-direct English LaTeX → out-of-tree build. Codifies non-negotiables (no em/en-dashes, no forward refs in background, citation content verified, scienceplots conventions, build never inside sync folder) and orchestrates overleap, scienceplot-py, reference-search, bibtex-gen, and commit-triage in turn.	overleaf-section-workflow/SKILL.md	TeX distribution (pdflatex, bibtex) + the companion skills it orchestrates
proton-mail	Read and search Proton Mail through a locally running Proton Bridge (read-only IMAP over STARTTLS on 127.0.0.1); does not send, delete, or move messages	proton-mail/SKILL.md	Proton Bridge running locally + ~/.proton-imap credentials file (chmod 600)
reference-search	Search and curate academic references via domain-aware routing across InspireHEP, OpenAlex, and Semantic Scholar for reports, claims, and section-level citation support	reference-search/SKILL.md	None (stdlib Python; optional S2_API_KEY env var)
research-backup	Mirror untracked research report directories (outputs/, results/, report(s)/) into the locally synced Dropbox folder with rsync, preserving the <category>/<project> layout so same-named projects never collide. discover.sh scans for candidate directories git does not track (ignored / untracked / outside any repo) and a registry under ~/.config/research-backup/ keeps the backup set explicit; backups are additive (no --delete), so local deletions never propagate.	research-backup/SKILL.md	rsync + the official Dropbox client syncing ~/Dropbox locally
research-log	Register projects and record decisions; surfaces the lesson/rule corpus at decision time (check step warns against past anti-patterns) and on cross-project recall when stuck	research-log/SKILL.md	~/.research/ workspace (auto)
research-portal	Build and manage a local MkDocs Material research portal over a folder of Typora/markdown notes: project-grouped sidebar via mkdocs-literate-nav, full LaTeX rendering via MathJax 3, and safe project tagging/renaming that repairs image links	research-portal/SKILL.md	uv; mkdocs-material and mkdocs-literate-nav are auto-installed by the scaffold script
research-report	Create or revise understanding-first research and experiment reports (why it matters, the core idea, what it achieved) backed by traceable evidence, plots, manifests, and validation helpers	research-report/SKILL.md	None (stdlib Python)
scienceplot-py	Generate a Python matplotlib plot script following the user's mandatory scienceplots (science+nature) lab template, with parquet / CSV / NumPy data sources and four plot variants (single line, multi-line, scatter/errorbar, subplots). Writes the .py only — does not execute it.	scienceplot-py/SKILL.md	matplotlib, scienceplots, plus pandas / numpy as needed (in the user's runtime env)
vastai	Search, create, and manage Vast.ai GPU cloud instances	vastai/SKILL.md	vastai CLI + API key
wide-slide-illustrator	Compose detailed image-generation prompts (ChatGPT Image 2.0, DALL-E, Sora, Midjourney) for wide cinematic 18:9 multi-panel infographic slides, six style variants: Friendly Whiteboard, Editorial Magazine, Engineering Blueprint, Swiss Minimalist, Dark Tech / Neon, Scientific Poster	wide-slide-illustrator/SKILL.md	None
workshop-paper-review	Produce OpenReview-ready ~400-word peer reviews for ICML/NeurIPS/ICLR workshop papers (4-8 page submissions): PDF intake, evidence-grounded drafting, anti-anchoring scoring calibration, parallel fact-check against the source PDF, AI-writing-pattern removal, and a Korean-draft to English-submission workflow	workshop-paper-review/SKILL.md	None (PDF reading; stdlib)
xkcd-py	Generate a Python matplotlib plot script following the user's mandatory xkcd lab template (with plt.xkcd():, figsize=(10, 6), dpi=300), with parquet / CSV / NumPy data sources and four plot variants (single line, multi-line, scatter/errorbar, subplots). Writes the .py only — does not execute it.	xkcd-py/SKILL.md	matplotlib, plus pandas / numpy as needed (in the user's runtime env)
zai-web-search	Search the live web via z.ai's web_search_prime MCP server (included with the GLM Coding Plan, so no separate API recharge): a single stdlib-Python3 script (scripts/web_search.py) does the MCP streamable-HTTP handshake, SSE parsing, and multiply-escaped-JSON handling, reading the z.ai key at runtime from pi's ~/.pi/agent/auth.json (single source of truth — the same key pi uses for the default model). Returns title / link / snippet per result, retries transient HTTP/TLS failures, and locally enforces the optional --domain restriction; --json is available for piping. Complementary to reference-search (InspireHEP / OpenAlex / Semantic Scholar): use this for recent trends, news, blogs, docs, and anything outside academic databases.	zai-web-search/SKILL.md	None (stdlib Python3; z.ai key from ~/.pi/agent/auth.json via GLM Coding Plan login)
zoom-summary	Read Zoom AI Companion meeting summaries through the Zoom REST API v2 (Server-to-Server OAuth): list summaries over a date range with full pagination, fetch a summary body, and save it as markdown. Also repairs the domain jargon the summariser mistranscribes via a two-tier glossary — high-confidence terms are substituted, low-confidence ones are only reported for a human to resolve — keeping the untouched summary as <stem>.raw.md so runs are idempotent and auditable.	zoom-summary/SKILL.md	Zoom Server-to-Server OAuth app with the two meeting-summary scopes, on a paid account with AI Companion enabled

Quick start

1. Install the skills you want with your preferred client — see CLIENT_SETUP.md for Claude Code, Codex, and Forge instructions.
2. Complete any external setup required by the skill (see the next section) before using it.
3. Open that skill's SKILL.md first for workflow details.
4. Use bundled scripts/ for executable helpers and references/ for deeper guidance.

Skill requirements & setup

Several skills depend on external CLIs, API keys, or credential files. Install and configure them once per machine before invoking the skill.

academic-jobs

Requires uv on PATH. The skill bundles a small uv project (ajo); uv run --project <skill-dir> ajo … auto-installs its dependencies (requests, beautifulsoup4) into an isolated environment on first run. No API keys.

• Searches two boards: AJO (academicjobsonline.org, HTML scraping) and InspireHEP (inspirehep.net/jobs, public JSON API). ajo fetch hits both by default; pick boards per preset (--sources ajo,inspire) or per run (--source ajo|inspire|both).
• State lives under ~/.local/share/academic-jobs/ (jobs.db + config.toml); override with the AJO_DATA_DIR env var. First run creates the data dir, the SQLite DB, and a default physics-ml preset. Postings are keyed by (source, id) (the boards use overlapping integer ids); a pre-InspireHEP v1 DB is migrated in place on first open.
• AJO validity is judged from each posting's detail page (effective deadline = firm Appl Deadline if present, else the listed until date), not the list page; --fast skips AJO detail pages but misses many valid postings. InspireHEP is queried with status=open and uses the structured deadline_date directly (no detail fetch, --fast does not apply).
• The CLI uses one polite session per board with a real User-Agent, a small inter-request delay, and a per-run AJO detail-fetch cap (reported under stats.per_source). Do not parallelise or hammer either board.
• See academic-jobs/references/: fetch.md, presets.md, schema.md.

adversarial-review

No external setup required. Spawns persona subagents through the host client's Agent tool and uses reference-search (stdlib Python + InspireHEP/OpenAlex/Semantic Scholar) for citation and prior-art audits.

bibtex-gen

No external setup required. The HEP (InspireHEP), publisher-fallback (CrossRef), and arXiv-category-probe paths are stdlib-only. The non-HEP Google Scholar path uses the scholarly library, which is declared in the orchestrator's PEP 723 inline script metadata header — uv run bibtex-gen/scripts/bibtex_gen.py … will provision an ephemeral environment that includes it on first run and reuse the cached env afterwards. No uv add or pip install needed.

• If the orchestrator is run with bare python3 instead of uv run, scholarly may be missing; non-HEP queries then fall through directly to CrossRef DOI bibtex (correct publisher-grade entries, but you lose Scholar's preferred key style).
• HEP / non-HEP classification is automatic — a query is HEP iff InspireHEP returns a match. arXiv IDs additionally consult the arXiv API for hep-* / nucl-* category tags. Use --hep or --no-hep to override.
• Source-native bibtex keys are preserved verbatim: Author:YYYYabc (InspireHEP), firstauthorYYYYword (Scholar), Lastname_Year (CrossRef). The skill does not rewrite keys.
• The orchestrator sleeps 0.5 s between batch queries by default to stay polite with all three APIs (--sleep 0 to disable).
• Helper script: bibtex-gen/scripts/bibtex_gen.py. See bibtex-gen/references/examples.md for full CLI patterns.

commit-triage

No external setup required. Uses only git in the current repository.

concept-explainer

Requires matplotlib and scienceplots in the runtime that executes the generated plot scripts, plus a working system TeX install (LaTeX rendering is mandatory — the no-latex style is forbidden by this skill). The md2pdf-typora skill handles the PDF assembly step at the end.

• Install runtime deps:

  uv add matplotlib scienceplots numpy

• The import scienceplots line in every generated plot script is required even though Pyright flags it as unused — the science and nature styles register by import side-effect.
• Unlike scienceplot-py, this skill does execute the plot scripts it generates (via uv run), because the auto-PDF step at the end depends on the PNGs existing.
• Optional: wide-slide-illustrator for Friendly Whiteboard schematic prompts (the skill outputs prompts; the user — or the codex-image skill — renders them).
• Every finished explanation is archived to ~/Dropbox/ConceptExplainer/<Topic>/YYYYMMDD_<concept-slug>/, reusing the local folder name so the topic listing sorts chronologically. Skip this if you do not want Dropbox mirroring.
• See concept-explainer/references/: audience-calibration.md, rigor-checklist.md, visualization-playbook.md, structure-template.md, schematic_friendly.md, plus plot_skeletons/{function_plot,parametric_sweep,heatmap}.py.

dropbox

Requires a Dropbox app and OAuth credentials stored at ~/.config/dropbox-skill/credentials.json.

• Dependencies: curl, jq.
• Run the interactive setup flow:

  bash dropbox/scripts/setup.sh

  The script prompts for your app key, app secret, and an authorization code from the Dropbox OAuth URL. Required app permissions: files.content.write, files.content.read, sharing.write, sharing.read.

hep-rumor-mill

Requires uv on PATH. The skill bundles a small uv project (prm); uv run --project <skill-dir> prm … auto-installs its one dependency (requests) into an isolated environment on first run. No API keys.

• Public, no-auth data sources: the rumor-mill Google Sheet (CSV export), the InspireHEP author + literature REST API (primary record, keyed off the recid the sheet links per person), OpenAlex (cross-disciplinary augmentation by ORCID, for everyone), the ORCID public API (author-claimed work list, used to rescue people whose OpenAlex ORCID cluster is conflated), and Semantic Scholar (name-based fallback when a person has no ORCID).
• State lives under ~/.local/share/hep-rumor-mill/rumor.db; override with the PRM_DATA_DIR env var. prm fetch stores the year's entries, prm enrich resolves records and is resumable (--limit N caps a run, re-run to continue).
• Two data-quality guards are built in and surfaced, not hidden: large-collaboration inflation (n_large_collab, InspireHEP papers with > 50 authors) and OpenAlex author-conflation (a shared-name ORCID cluster is dropped when contaminated with distant-field works). The rumor mill is self-reported and incomplete; the analysis is descriptive, never predictive.
• The CLI paces InspireHEP requests and OpenAlex paging; Semantic Scholar rate-limits hard and is used for the no-ORCID author fallback and arXiv-only citation backfill. Optional: set S2_API_KEY (same env var reference-search uses) to raise the Semantic Scholar limit and make citation backfill reliable; unauthenticated works but is best-effort. Do not parallelise or hammer the APIs.
• See hep-rumor-mill/references/: sheets.md, metrics.md, analysis.md, schema.md.

wide-slide-illustrator

No external setup required. The skill is purely a prompt composer — it produces a copy-pasteable text prompt for an image-generation model (ChatGPT Image 2.0 / gpt-image-1, DALL-E 3, Sora image, Midjourney). You bring your own image generator.

• Six reusable style variants — all share the same 5-panel 18:9 composition and English-only-on-canvas guard rail; they differ only in surface treatment.
  • Friendly Whiteboard (warm sketch, lab-meeting feel): wide-slide-illustrator/references/friendly-whiteboard-style.md; example example-osprey-v021.md.
  • Editorial Magazine (Quanta / NYT feature spread): wide-slide-illustrator/references/editorial-magazine-style.md; example example-osprey-editorial.md. Sub-patterns: STATUS-BADGE / PULL-QUOTE / CONNECTION-LINE.
  • Engineering Blueprint (NASA / SpaceX schematic, monospace + cyan-on-navy): wide-slide-illustrator/references/engineering-blueprint-style.md. Sub-patterns: DIMENSION-LINE / TERMINAL-NODE / STAGE-BOX.
  • Swiss Minimalist (Müller-Brockmann typography over methodology content; orthodox title-poster mode opt-in via tone dial): wide-slide-illustrator/references/swiss-minimalist-style.md. Sub-patterns: PROMINENT-NUMBER / PRIMARY-PLOT / GRID-CARD.
  • Dark Tech / Neon (Linear / Anthropic / Cursor keynote, charcoal + neon glow): wide-slide-illustrator/references/dark-tech-neon-style.md. Sub-patterns: NEON-CHIP / GLOW-PATH / MONOSPACE-LABEL.
  • Scientific Poster (Phys Rev / Nature published figure, panel labels (a)–(e), running FIG. N. caption): wide-slide-illustrator/references/scientific-poster-style.md. Sub-patterns: PANEL-LETTER / FIG-CAPTION / AXIS-LABEL.
• Style blocks are load-bearing — drop them into prompts verbatim, do not paraphrase. Always include the variant's "hard negatives" bullet so cues from sibling variants don't leak through (lesson learned from the editorial v1 → v2 hardening pass).
• Defaults: 18:9 cinematic-wide canvas, English-only on-canvas text, numbered marker per panel. See SKILL.md for tone dials per variant (calmer / more playful / Bauhaus-clean / orthodox-Swiss-poster / Korean-caption hybrid, etc.).

handdrawn-schematic

Renders a single friendly hand-drawn whiteboard schematic on a pure white background (no cream/beige tint) that explains a schema (concept, pipeline, architecture, algorithm) at a glance. By default it produces a PNG via the bundled codex image_generation tool (ChatGPT OAuth, no API key); without codex it falls back to emitting the finished prompt for any image generator.

• The style block in handdrawn-schematic/references/style-block.md is load-bearing — drop it into the prompt verbatim, do not paraphrase. It is the journal-club friendly-whiteboard block with the background changed to pure white (#FFFFFF).
• Build a 3-6 panel figure brief (headline, subtitle, panels flowing left to right); panels are English-only, labels <= 6 words, charts stay hand-drawn.
• This is the reusable single-figure generator extracted from journal-club-review. For the paired method+results figures inside a paper walkthrough use that skill; for a prompt-only composer across six other styles (editorial, blueprint, swiss, neon, poster) use wide-slide-illustrator; for matplotlib data plots use scienceplot-py / xkcd-py.

journal-club-review

Requires uv (the extractor declares its deps inline via PEP 723: pdfplumber, httpx, feedparser). Figures need a logged-in bundled codex runtime (ChatGPT OAuth); without it the review still renders text-only.

• Ingest an arXiv id/URL, a PDF, or a text/markdown file with scripts/extract_text.py (writes source.md + JSON metadata to reviews/<slug>/).
• Output language auto-matches the source paper unless the user asks otherwise; all math is rendered as LaTeX.
• Two optional friendly-whiteboard infographics (method + results) are generated in parallel via codex exec.
• This is a teaching/discussion artifact (no scores, no accept/reject). For a referee report use workshop-paper-review; for an adversarial audit use adversarial-review.

md2pdf-typora

Requires pandoc and a Chromium-family browser (google-chrome-stable, google-chrome, or chromium) on PATH. The Whitey theme CSS bundled at md2pdf-typora/typora-whitey.css is used inline; no separate install.

• Install pandoc with your package manager:

  # Arch
  sudo pacman -S pandoc
  # Debian / Ubuntu
  sudo apt install pandoc
  # macOS
  brew install pandoc

• A working Chromium build is required because the pipeline renders the patched HTML to PDF via --headless=new --print-to-pdf.
• Math is rendered via MathJax SVG (tex-svg-full.js) loaded from CDN — Chrome needs network access during the conversion. The skill rewrites pandoc's default CHTML reference to SVG to avoid silent missing-glyph fallbacks (e.g. \phi rendering as □).
• Body fonts (IBM Plex Serif Latin, MaruBuri Korean, Roboto Slab headings, JetBrains Mono code) are also fetched from Google Fonts CDN.
• Optional flags: --dropbox [subfolder] mirrors the produced PDF to ~/Dropbox/Magi/[subfolder]/; --send-telegram posts it via the Telegram skill if configured.

morgen

Requires a Morgen API key stored at ~/.config/morgen-skill/credentials.json.

• Dependencies: curl, jq.
• Get an API key at https://platform.morgen.so under Developers API.
• Run the interactive setup:

  bash morgen/scripts/setup.sh

  The script prompts for the API key, verifies it against /v3/calendars/list, and writes the credentials file with mode 600.
• All requests go to https://api.morgen.so/v3 with the header Authorization: ApiKey <KEY>.
• Rate limit: 100 points per 15 minutes (list endpoints cost 10 points each).

overleap

Requires the overleap Node.js CLI and a valid Overleaf session cookie.

• Install:

  npm install -g overleap

  Requires Node.js >= 18 and git (the dependency socket.io-client is fetched from a GitHub fork).
• Verify the binary is on PATH:

  command -v overleap && overleap --help | head -5

• Save your Overleaf session cookie into a per-project .env file with the bundled helper:

  bash overleap/scripts/init_env.sh <dir> <<< 'overleaf_session2=...'

  The script writes <dir>/.env with mode 0600 and ensures .env is in <dir>/.gitignore. Get the cookie from browser DevTools → Application → Cookies → overleaf.com → copy the full cookie string (or just the overleaf_session2 value). See overleap/references/cookie-setup.md for full instructions and refresh procedure.
• Smoke-test:

  cd <dir> && overleap projects

• overleap sync is a long-running daemon — never invoke it as a foreground Bash call from Claude. Use a separate terminal, pueue, or run_in_background (overleap/references/long-running-patterns.md covers all three).

overleaf-section-workflow

Requires a working TeX distribution on PATH and the companion skills it orchestrates (all already in this repo).

• TeX prerequisites (used by the templates/build_template.sh build script):

  command -v pdflatex && command -v bibtex

  Install via TeX Live (sudo pacman -S texlive-most on Arch, sudo apt install texlive-full on Debian/Ubuntu, brew install --cask mactex on macOS) — the skill does not install TeX for you.
• Companion skills the workflow invokes (no extra setup beyond their own per-skill prerequisites): overleap (Overleaf sync), scienceplot-py (plot scripts), reference-search (literature discovery), bibtex-gen (citation generation), commit-triage (clean session-boundary commits), and optionally deep-research for novelty verification.
• Three-directory contract: the workflow assumes ~/zbin/OverLeaf/<PROJECT>/ (sync), <PROJECT>_draft/ (Korean drafts), <PROJECT>_build/ (out-of-tree TeX build). Never build inside the sync folder — build artefacts would propagate back to Overleaf.
• Translation is Opus-only by user mandate — this skill explicitly forbids delegating section translation to a Sonnet subagent, even where the general lab convention would. See references/05_translation_rules.md.
• All draft text and final LaTeX must be em/en-dash-free; verify with grep -cP "[\x{2013}\x{2014}]" <file> (must return 0).

proton-mail

Requires a locally running Proton Bridge instance and a credentials file at ~/.proton-imap (chmod 600).

• The ~/.proton-imap file holds KEY=VALUE lines: PROTON_IMAP_USER, PROTON_IMAP_PASS (the bridge-specific password from the Bridge app, not your main Proton account password), PROTON_IMAP_HOST=127.0.0.1, and PROTON_IMAP_PORT=1143.
• The skill connects read-only over STARTTLS on 127.0.0.1 at the port Proton Bridge exposes. It does not send, delete, or move messages.
• Start Proton Bridge before invoking the skill; the bridge must remain running for the duration of the session.

reference-search

No external setup required. Uses Python standard library only and queries InspireHEP, OpenAlex, and Semantic Scholar via their public APIs.

• Domain is inferred automatically from the query; HEP-family domains route to InspireHEP first, others to OpenAlex with Semantic Scholar as fallback. Override with --domain or --source.
• Optional: pass --email for OpenAlex's polite pool.
• Optional: set S2_API_KEY env var to raise Semantic Scholar rate limits (unauthenticated works but is rate-limited).
• Helper script: reference-search/scripts/reference_search.py.

research-backup

Requires rsync and the official Dropbox client syncing a local folder (default destination: ~/Dropbox/ResearchBackup/). No API keys: the skill only writes files under the synced folder and the daemon uploads them.

• Config and registry live under ~/.config/research-backup/ (config + registry), both auto-created with defaults on first run; override the directory with RESEARCH_BACKUP_CONFIG_DIR.
• Defaults: SCAN_ROOT="$HOME/Documents/Project", BACKUP_ROOT="$HOME/Dropbox/ResearchBackup", DIR_NAMES="outputs results report reports". Add rsync exclude patterns via RSYNC_EXCLUDES (e.g. "*.ckpt wandb").
• Workflow: discover.sh scans and classifies candidates (NEW / REGISTERED / TRACKED / MISSING; git-tracked directories are never registered), discover.sh --register appends the NEW ones, backup.sh --dry-run --all previews, backup.sh --all (or a path filter like backup.sh MyProject) syncs.
• Backups are additive by design: no --delete is ever passed, so files removed locally remain in Dropbox.

research-log

No external setup required. The skill creates and manages ~/.research/ on first use. The redesign shifts focus from a write-heavy diary to an advisor: a check step surfaces relevant past lessons and rules before a decision, and a recall step surfaces cross-project findings when you are stuck. The workspace still holds per-project decision entries and session state.

• Workspace: ~/.research/ (auto-created on first use).
• See research-log/references/conventions.md for the storage layout.

research-portal

Requires uv on PATH. The mkdocs dependencies (mkdocs-material and mkdocs-literate-nav) are auto-installed by the skill's scaffold script into an isolated environment; no manual pip install is needed.

• docs_dir points directly at your notes folder, so original files are never modified.
• LaTeX is rendered via pymdownx.arithmatex + MathJax 3.
• Project tagging and renaming operations repair image links automatically.
• See research-portal/SKILL.md for the full scaffold and serve workflow.

research-report

No external setup required. Uses Python standard library only.

• Helper scripts live in research-report/scripts/ (build_plot_manifest.py, record_report_version.py, validate_artifacts.py).
• Designed to be used from inside a project directory that already contains experiment outputs.

scienceplot-py

Requires matplotlib and scienceplots in the runtime environment that will execute the generated script. The skill writes the .py file but does not run it; the user runs it themselves (preferred: uv run <path>).

• Install runtime deps in the project that owns the data:

  uv add matplotlib scienceplots pandas pyarrow numpy

  • pandas for parquet / CSV input. pyarrow is the parquet backend (drop it for CSV-only or NumPy-only data).
  • numpy for .npy / .npz input.
• The import scienceplots line in the generated script is required even though Pyright flags it as unused — the science and nature styles register by import side-effect.
• savefig DPI is 300 (intentionally lowered from the upstream pq_plot.py template's 600). Bump higher only if the user explicitly asks.
• Plot variants and data-loader patterns: see scienceplot-py/references/ (single_line.py, multi_line.py, scatter_errorbar.py, subplots.py, data_loaders.md).

vastai

Requires the vastai CLI and a Vast.ai API key.

• Install the CLI:

  uv pip install vastai   # or: pip install vastai

• Get your API key from https://cloud.vast.ai/cli/ and register it:

  vastai set api-key YOUR_API_KEY

  The key is stored at ~/.vast_api_key. Never commit or share it.
• Verify access: vastai show user.

workshop-paper-review

No external setup required. Uses PDF reading and Python standard library only; no API keys or CLIs needed.

• The skill reads the source PDF to fact-check claims in the review draft, using only the local file.
• AI-writing-pattern removal and anti-anchoring score calibration are built into the drafting workflow.
• See workshop-paper-review/SKILL.md for the full review workflow and OpenReview submission checklist.

xkcd-py

Requires matplotlib in the runtime environment. The skill writes the .py file but does not run it; the user runs it themselves.

• Install runtime deps:

  uv add matplotlib pandas pyarrow numpy

  No scienceplots is needed — plt.xkcd() is built into matplotlib.
• For best visual results, install an xkcd-style font (Humor Sans / xkcd Script / Comic Neue). Without it, matplotlib falls back to Bitstream Vera Sans and emits a findfont: Font family ['xkcd Script', ...] not found warning; the plot still renders correctly.
• savefig DPI is 300 with figsize=(10, 6) — a wider canvas than the matplotlib default, needed for the hand-drawn font and stroke widths to read cleanly.
• Plot variants and data-loader patterns: see xkcd-py/references/ (same four templates as scienceplot-py, plus data_loaders.md).

zai-web-search

Requires the z.ai key under zai-coding-cn.key in ~/.pi/agent/auth.json — the same key pi already uses for the default model, included with the GLM Coding Plan login. No pip install: scripts/web_search.py is stdlib Python3 and handles the MCP streamable-HTTP handshake, SSE parsing, and multiply-escaped JSON inline.

• Verify the key is present:

  test -f ~/.pi/agent/auth.json && python3 -c "import json;print(json.load(open('/home/axect/.pi/agent/auth.json'))['zai-coding-cn']['key'])" >/dev/null && echo OK

• The MCP endpoint (web_search_prime) is covered by the Coding Plan — no separate API recharge. Do not duplicate the key into a second file; auth.json is the single source of truth.
• --domain example.com is checked locally against result hostnames because the upstream filter can return off-domain results; the requested domain and its subdomains are accepted.
• Transient HTTP/TLS connection failures are retried twice before returning exit code 5.
• Distinct from reference-search: this skill searches the live web (news, blogs, docs, repos, trends). Use reference-search for academic papers, metadata, and BibTeX, and run both when a claim needs a primary source plus current context.

zoom-summary

Requires a Zoom Server-to-Server OAuth app on a paid account with AI Companion meeting summaries enabled, created by the account owner at https://marketplace.zoom.us (Develop → Build App). Dependencies: curl, jq, base64, plus python3 for correct.py.

• Add both scopes on the app's Scopes tab: meeting:read:list_summaries:admin (list) and meeting:read:summary:admin (body). If the picker does not offer the granular pair, the classic meeting_summary:read:admin covers both. Activate the app, then register the credentials:

  bash zoom-summary/scripts/setup.sh --stdin <<< '{"account_id":"...","client_id":"...","client_secret":"..."}'

  setup.sh reports GRANTED/MISSING per scope from the token's own scope string and then probes the list endpoint live, so a half-provisioned app is visible immediately rather than at first use.
• The body scope is reported missing from the scope picker on some accounts (open Zoom developer-forum reports as of 2026-05). The symptom is that list works and get returns HTTP 403. There is no API workaround; the zoom.us web portal (Meeting Summary with AI Companion → My Summaries) is the fallback.
• Credentials live at ~/.config/zoom-skill/credentials.json (mode 600); tokens are cached for their one-hour life at ~/.cache/zoom-skill/token.json, keyed by a fingerprint of the credentials.
• references/glossary.tsv is a template that ships with no real terms. Every real entry goes in ~/.config/zoom-skill/glossary.tsv, which is outside the repository, so project vocabulary and personal names are never published here. Back that file up separately; research-backup covers it as an _external/ entry.

Which skill to use?

• Choose academic-jobs to pull current, still-open academic job postings (postdoc / faculty / PhD) from Academic Jobs Online and the InspireHEP jobs board (searched together by default), filtered to postings whose application deadline has not passed, with field presets and "new since last check" tracking.
• Choose adversarial-review to stress-test a paper draft or report before submission, simulate hostile referees, or audit citations and figures.
• Choose bibtex-gen to build a .bib file or one-off bibtex entries from arXiv IDs / DOIs / paper titles — HEP papers are routed to InspireHEP, non-HEP papers go to Google Scholar with CrossRef DOI bibtex as the publisher fallback, and source-native keys are preserved verbatim.
• Choose commit-triage to tidy a noisy working tree, archive failed experiments to failure/, and produce clean grouped commits.
• Choose concept-explainer to write a kind-but-rigorous explanation of one concept for a named audience — explanation.md with full derivations (every symbol defined, every step's rule named, =/≈/∼ disciplined), executable scienceplots ["science", "nature"] matplotlib plots (no-latex forbidden), optional Friendly Whiteboard schematics for "the big picture", and an auto-rendered PDF (archived to Dropbox under a date-prefixed folder).
• Choose dropbox for file upload, download, or shared-link workflows in Dropbox.
• Choose hep-rumor-mill to study the HEP-theory postdoc job market from the rumor mill: who got offers at which institutions, the publication profiles of the people who got them (citations, papers, venues, subfield, PhD-age, named fellowships, with OpenAlex cross-disciplinary augmentation for interdisciplinary candidates), and how your own InspireHEP profile sits against the accepted cohort. Self-reported data, descriptive not predictive.
• Choose handdrawn-schematic to generate a single friendly hand-drawn whiteboard schematic on a pure white background (wavy marker strokes, numbered panels left-to-right, chunky chalk arrows, hand-written notes) that explains a concept, pipeline, architecture, or algorithm at a glance. Renders a PNG by default via the bundled codex image_generation tool, or emits the copy-paste prompt when codex is unavailable. Use wide-slide-illustrator instead for a prompt-only composer across six other styles, journal-club-review for the paired method+results figures inside a paper walkthrough, and scienceplot-py / xkcd-py for real matplotlib data plots.
• Choose journal-club-review to turn an arXiv id/URL, a PDF, or raw text into a journal-club presentation (nine sections from TL;DR to Takeaways) meant to help a reading group understand and discuss a paper, with LaTeX math, source-language auto-matching, and two optional friendly-whiteboard figures. Use workshop-paper-review instead for an OpenReview referee report, or adversarial-review for a hostile pre-submission audit.
• Choose md2pdf-typora to convert a Markdown report or note (especially one with LaTeX math, Korean text, and embedded plots) into a print-ready PDF that visually matches Typora's Whitey-theme export.
• Choose morgen for calendar and task management across accounts connected to Morgen (Google, Microsoft 365, iCloud, Fastmail, CalDAV) and native Morgen tasks/tags.
• Choose overleap to edit Overleaf projects locally with real-time bidirectional sync — local edits propagate to Overleaf and vice versa, so Claude can edit .tex files and collaborators see them on Overleaf instantly.
• Choose overleaf-section-workflow when you are drafting a physics paper section-by-section on Overleaf and want the disciplined Korean-draft → user-iteration → Opus-direct English-LaTeX → out-of-tree-build loop, with citation-content verification, scienceplots-grade plots, and em/en-dash-free output. This is the orchestration layer; overleap is just the sync primitive it builds on.
• Choose proton-mail to search or read Proton Mail messages and threads through a locally running Proton Bridge, useful for retrieving paper notifications, calendar invites, or collaboration emails without leaving the terminal.
• Choose reference-search for literature search, citation curation, and section-level reference support when drafting reports.
• Choose research-backup to back up the untracked outputs/ / results/ / report(s)/ directories of your research projects into the locally synced Dropbox folder, organized as <category>/<project> so nothing collides. Additive rsync mirroring, no API calls. Choose dropbox instead for one-off file upload/download/share through the API.
• Choose research-log to register projects, record decisions, check past lessons before committing to a new approach (the check step warns against known anti-patterns), or recall cross-project findings when stuck (the recall step).
• Choose research-portal to build a browsable, math-rendered MkDocs Material site over a folder of existing Typora/markdown notes without modifying the originals.
• Choose research-report to turn research or experiment outputs into an understanding-first report: one that explains why the work matters, what the core idea is (intuition before formalism), and what it achieved, with every claim earned by traceable evidence and plots.
• Choose scienceplot-py to scaffold a publication-style matplotlib script following the lab's scienceplots (science+nature) template — single line, multi-line + legend, scatter / errorbar, or multi-panel subplots from parquet / CSV / NumPy data.
• Choose vastai for renting GPUs, searching offers, or managing remote compute instances.
• Choose wide-slide-illustrator to compose image-generation prompts for wide cinematic 18:9 multi-panel infographic slides: pipeline diagrams, "how it works" figures, hero figures, paper figures, keynote backdrops. Six style variants by audience and target medium: Friendly Whiteboard (lab meeting), Editorial Magazine (Quanta / NYT paper hero), Engineering Blueprint (technical schematic), Swiss Minimalist (typographic poster on methodology content), Dark Tech / Neon (AI-lab keynote backdrop), Scientific Poster (Phys Rev / Nature journal figure).
• Choose workshop-paper-review to write an OpenReview-ready peer review for a workshop paper submission (ICML/NeurIPS/ICLR, 4-8 pages): the skill handles PDF intake, evidence-grounded drafting, anti-anchoring score calibration, and AI-writing-pattern removal, with a Korean-draft to English-submission option.
• Choose xkcd-py for a hand-drawn / sketch-style matplotlib script (with plt.xkcd():, wider canvas, dpi=300) — same data-source and plot-variant coverage as scienceplot-py.
• Choose zai-web-search for live web search via z.ai (recent trends, news, blogs, docs, repos) when the need is clearly outside academic databases; the bundled script reads the key from pi's auth.json and the MCP endpoint is covered by the GLM Coding Plan. Complementary to reference-search — run both and merge when a claim needs a primary source plus current context.
• Choose zoom-summary to pull a past Zoom meeting's AI Companion summary into the session — list what exists over a date range, fetch one, save it as markdown, and repair the mistranscribed jargon before you act on it. Read-only: it cannot generate a summary for a meeting that did not have AI Companion on.

Structure

skills/
├── academic-jobs/
├── adversarial-review/
├── bibtex-gen/
├── commit-triage/
├── concept-explainer/
├── deprecated/              # retired skills, entrypoint renamed so nothing loads them
├── dropbox/
├── handdrawn-schematic/
├── hep-rumor-mill/
├── journal-club-review/
├── md2pdf-typora/
├── morgen/
├── overleap/
├── overleaf-section-workflow/
├── proton-mail/
├── reference-search/
├── research-backup/
├── research-log/
├── research-portal/
├── research-report/
├── scienceplot-py/
├── vastai/
├── wide-slide-illustrator/
├── workshop-paper-review/
├── xkcd-py/
├── zai-web-search/
└── zoom-summary/

Deprecated skills

Retired skills live under deprecated/, with the entrypoint renamed SKILL.md to SKILL.md.deprecated so no harness discovers them (Pi scans this repository root recursively). See deprecated/README.md for the current list, the reasoning, and how to restore one.

License

Released under the MIT License © 2026 Tae-Geun Kim.