Claude Code Skills — QuantumBFS

A small marketplace of Claude Code plugins, each shipping one skill.

Install

In Claude Code, add this repo as a marketplace, then install whichever plugin(s) you want:

/plugin marketplace add QuantumBFS/claude-code-skills
/plugin install download-papers@quantum-bfs
/plugin install submit-slurm-job@quantum-bfs
/plugin install verify-references@quantum-bfs
/plugin install paper-review-checklist@quantum-bfs
/plugin install distill-feedback-from-history@quantum-bfs
/plugin install digitize-plots@quantum-bfs
/plugin install claude-log-restore@quantum-bfs

The @quantum-bfs suffix matches the name field in .claude-plugin/marketplace.json.

To update later:

/plugin marketplace update quantum-bfs

Plugins

download-papers

Download an academic paper as a PDF given a URL, DOI, paper title, or arXiv ID. Resolves in order: open web (author homepage, repositories) → Sci-Hub mirrors → arXiv.

Triggers on phrases like "download paper", "get me the PDF", "fetch this DOI". No configuration required.

submit-slurm-job

Generate and submit sbatch scripts for GPU compute jobs on a Slurm cluster. Handles partition selection, explicit GPU model in --gres, full Python path, and log file management.

Required configuration — add to your project's CLAUDE.md:

## Slurm Configuration

- PYTHON_PATH: `/path/to/miniconda3/envs/myenv/bin/python3`
- PROJECT_DIR: `/home/your_user/private/homefile`   # scripts/logs must live here
- PARTITION: `home`                                 # compute partition

Invoke with /submit-slurm-job (or just describe the job — Claude will pick up the skill).

verify-references

Verify a BibTeX bibliography against CrossRef metadata. The skill:

1. Scans your .tex files to flag uncited entries.
2. Queries CrossRef for every entry with a DOI and writes a side-by-side .bib.
3. Compares the two and produces a Markdown report listing field-level discrepancies (title, author, journal, volume, pages, year), sorted by similarity.

Useful before submission to catch hallucinated or stale BibTeX. Invoke by asking to "verify references" / "check the bib file". Requires Python 3 and an internet connection (for the CrossRef API).

paper-review-checklist

Review a LaTeX research-paper draft against a fixed checklist (notation, figures, citations, structure, equations, style, reproducibility) and insert inline author-comment reminders at the locations that fail each check.

The skill reads the entire .tex, runs programmatic checks (broken \ref/\cite, unreferenced equation labels, symbol-synonym pairs, one-use acronyms), then walks the checklist category by category and inserts short \lw{...} notes (or \todo{...} — whichever author-comment macro the project's preamble defines). Ends with a per-category summary and the top three most impactful issues.

Invoke by asking to "review my paper" / "apply the paper-review checklist to file X".

distill-feedback-from-history

Mine your own local Claude Code session history (~/.claude/projects/) for recurring behavioral feedback — corrections, format/voice edicts, error catches, domain-knowledge injections — and distill each pattern into the right kind of artifact:

• CLAUDE.md addition (default) — the natural home for discipline rules and preferences.
• Auto-memory feedback entry — alternative for the same content if you prefer the harness memory system.
• SKILL.md draft — only when the pattern describes a parameterizable multi-step workflow. Rare; behavioral patterns are not workflows.

The pipeline runs in eight phases: extract → tag (regex) → domain-injection (LLM) + filter → cluster → classify into a bucket → distill (autonomous or A/B/C/D/E gate) → render → hand back. Default scope is all projects, last 30 days.

Invoke by asking to "distill recurring feedback from my sessions" / "what should I add to CLAUDE.md?" / "find behavioral patterns I should write down". Single-user, local-only; reads JSONLs and writes drafts to ./distilled/, no network calls, no auto-install.

Note: This plugin replaces the earlier extract-skills-from-history, which mis-rendered behavioral rules as standalone SKILL.md files. The new name and bucketed output reflect that most patterns mined this way are memories, not skills.

digitize-plots

Extract numeric (x, y) data from a plot/figure image — including a figure inside a PDF — into a CSV. The command-line counterpart to WebPlotDigitizer, which is GUI-only.

The bundled digitize.py wraps the plotdigitizer CLI and hides its two silent footguns (it int-truncates fractional calibration values, and it expects bottom-origin pixel-y), then adds the surrounding workflow: render a PDF page to PNG, inspect to read calibration ticks off zoomed margins, and extract with optional single-series color isolation and a verification overlay. The whole job is usually two or three commands.

Invoke by asking to "digitize this plot", "extract the data/curve from this figure", "read the points off this chart", or "WebPlotDigitizer but from the CLI". Requires Python 3 with plotdigitizer, numpy, and Pillow (pip install plotdigitizer numpy Pillow); PDF rendering uses PyMuPDF or the pdftoppm binary.

claude-log-restore

Recover context from Claude Code session transcripts. Claude Code stores each session as a local JSONL file under ~/.claude/projects/<encoded-project-path>/<session-id>.jsonl; this skill locates the right one and mines it.

The bundled claude_log_restore.py is dependency-free and streams JSONL line by line, so it stays safe on multi-MB transcripts. It can list recent sessions for a project, find a session by partial UUID/hash (--all-projects), search for keywords (-q), or surface handoff/summary/next-step notes (--handoff), printing the transcript path, session id, an approximate claude --resume command, and selected snippets (--json for machine-readable output). It reads and summarizes only — in-session code/conversation restore is still done with /rewind inside Claude Code.

Invoke by asking to "restore my Claude session", "find the handoff notes from last time", "resume session 13e617", or "what were the next steps in this project's history?". Requires Python 3; no network calls, no dependencies.

Repository layout

.
├── .claude-plugin/
│   └── marketplace.json          # marketplace manifest, lists all plugins
└── plugins/
    ├── download-papers/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/download-papers/SKILL.md
    ├── submit-slurm-job/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/submit-slurm-job/SKILL.md
    ├── verify-references/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/verify-references/
    │       ├── SKILL.md
    │       ├── check_unused_refs.py
    │       ├── compare_refs.py
    │       └── download_crossref.py
    ├── paper-review-checklist/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/paper-review-checklist/SKILL.md
    ├── distill-feedback-from-history/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/distill-feedback-from-history/
    │       ├── SKILL.md
    │       └── scripts/
    │           ├── extract_turns.py
    │           ├── tag_turns.py
    │           └── write_artifact.py
    ├── digitize-plots/
    │   ├── .claude-plugin/plugin.json
    │   └── skills/digitize-plots/
    │       ├── SKILL.md
    │       └── scripts/digitize.py
    └── claude-log-restore/
        ├── .claude-plugin/plugin.json
        └── skills/claude-log-restore/
            ├── SKILL.md
            └── scripts/claude_log_restore.py

Helper scripts shipped alongside a SKILL.md are referenced from the skill instructions via ${CLAUDE_SKILL_DIR}, which Claude Code expands to the skill's install directory at invocation time.

Contributing

Add a new plugin by:

1. Create plugins/<your-plugin>/.claude-plugin/plugin.json (use one of the existing files as a template).
2. Add the skill at plugins/<your-plugin>/skills/<your-plugin>/SKILL.md with YAML frontmatter (name, description).
3. Append an entry to .claude-plugin/marketplace.json.

Keep skills well-scoped, parameterize any user-specific paths, and document required configuration in this README.

License

MIT