메뉴
Submit experiments or agent jobs with resource limits, check queue status, pause/resume workloads, troubleshoot slow machine
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
This skill should be used when the user asks to "commit and push", "commit push", "sync changes", "push changes", "commit and sync", or "update remote". Handles the full workflow of committing changes, pulling with rebase, and pushing to remote.
Sweep recent AI safety research from curated sources (Anthropic alignment science / red team, OpenAI, GDM, Apollo, Redwood, METR, FAR AI, Truthful AI, alphaxiv, arXiv) and surface items matching tracked topic terms (inoculation prompting, reward hacking, exploration hacking, metagaming, eval gaming, OOCR, scheming, alignment faking, sandbagging, etc.). Use when asked to "sweep AI safety", "what's new in alignment", "any recent papers on X", "weekly safety digest", or for staying current on AI safety literature.
Anthropic visual style for plots, diagrams, slides, and web. Use when creating any visual output that should have Anthropic's look-and-feel — matplotlib charts, TikZ diagrams, HTML/CSS, or presentations.
Catch LLM-fabricated citations in BibTeX files. Verifies arXiv/OpenReview entries against live metadata (titles, first authors), then guides manual verification of authorless prose claims. Use before submitting papers, after any LLM-assisted citation generation, or when a reference smells off.
Fact-check prose claims in slides, reports, PDFs, and papers — statistics, comparatives, attributions, causal claims, quotes. Two-pass extract-then-verify protocol with strict numerical precision and a doc-only mode. Use when the user asks to "check the claims in this deck", "fact-check this report", "audit this PDF", "verify the numbers in these slides", or before publishing/shipping any externally-facing document with quantitative claims. Complements `check-bib-references` (which handles BibTeX entries) — this skill handles the prose around them.
Log a one-line knowledge gap to the project's gaps.md file. Use when the user is surprised by Claude's answer, says "I didn't know that", "wait what", or wants to record a misconception they just discovered. Format "I assumed X but actually Y". Personal misconception log — much higher learning signal than feedback memories.
| name | jobs |
| description | Submit experiments or agent jobs with resource limits, check queue status, pause/resume workloads, troubleshoot slow machine |
Local job queue with cgroup-enforced CPU/memory limits.
| Command | What it does |
|---|---|
jexp <cmd> | Submit experiment (resource-capped) |
jagent <cmd> | Submit agent job (resource-capped) |
jclaude <args> | Headless claude --print through agent queue |
jls | Queue status |
jlog [id] | Job output |
jfollow <id> | Stream live output |
jpause <group|all> | Pause group |
jresume <group|all> | Resume group |
jkill <id> | Kill job |
jclean | Remove completed |
jwatch | Live dashboard |
jtop | Status + cgroup usage |
jguard | Memory pressure check |
| Group | Use | Parallel | CPU | Memory |
|---|---|---|---|---|
| experiments | ML training, heavy compute | 1 | per resources.conf | per resources.conf |
| agents | Claude Code, Codex CLI | 3 | per resources.conf | per resources.conf |
Run an experiment: jexp python train.py --epochs 100
Machine feels slow: jguard to check pressure, jpause experiments to free resources
Multiple agents: jagent claude --print "review src/" — queued if slots full
Scale up: Edit config/resources.conf, run ./deploy.sh --pueue
| Problem | Fix |
|---|---|
systemd --user not available | loginctl enable-linger $(whoami) |
| Memory limits silently ignored | sudo systemctl set-property user-$(id -u).slice Delegate=yes |
| pueued not running | systemctl --user start pueued |
