Skill Codex
Run a short, bounded autonomous loop of debugging or experiments before checking back in. Use when the user wants the agent to iterate for a few cycles, keep going autonomously, or stop only after a fixed budget or explicit stop rule is met.
Skill Codex
Design the smallest useful experiment matrix that separates important hypotheses and changes a decision. Use when the agent is proposing runs, ablations, sweeps, benchmarks, or debugging experiments across research or engineering projects.
Skill Codex
Create a short-lived, restart-ready handoff so a fresh agent can resume work without rereading the full conversation, and delete it once read. Use when pausing work, ending a session, switching agents, preparing a clean-context restart, or finding a handoff addressed to you.
Skill Codex
Directly monitor long-running jobs or pipelines, classify them as healthy, waiting, stalled, failed, completed, or unknown, and choose the next action. Use when the active agent needs to babysit training runs, batch jobs, services, experiments, or scheduled work over time.
Skill Codex
Record work in a lab notebook so it is both a literal activity log and an efficient re-entry aid. Use when the agent is running experiments, debugging workflows, triaging jobs, or updating project state that must be recoverable later.
Skill Codex
Run safe, high-yield codebase maintenance passes that trim dead code, vestiges, duplicate implementations, stale scripts/configs, and excessive LOC without churn. Use when the agent is asked for a maintenance pass, cleanup pass, code cleanup, slop trimming, dead-code removal, de-duplication, vestige pruning, or…
Skill Codex
Coordinate multiple long-running agents or workers while preserving the global goal, current evidence, and timely steering. Use when supervising tmux or managed agents, delegating parallel work, reconciling queued corrections, preventing workers from acting on stale direction, or serving as mission control across…
Skill Codex
Create, edit, render, preview, and export Quarto reveal.js slide decks for scientific and technical presentations. Use when the agent works on .qmd slides, reveal.js layout, speaker notes, figure sidecars, slide PDF export, or reusable deck templates.
Skill Codex
Detect when work is stuck in a low-yield loop and force a better reset move. Use when the agent is repeating the same kind of fix, relaunch, sweep, or analysis without materially improving understanding or changing the decision.
Skill Codex
Check whether findings make sense before trusting or reporting them by comparing to the nearest reference, asking what mechanism could explain them, and naming a falsifier. Use when the agent is interpreting results, summarizing progress, assessing regressions, or deciding whether an observation should change the plan.
Skill Codex
Review a claim, result, or next-step recommendation like a sharp labmate who is trying to catch overclaiming, bad baselines, confounds, and missing checks. Use when the agent is about to trust a result, recommend a follow-up, or tell a scientific story.
Skill Codex
Inspect queue state, submit or cancel jobs, debug sbatch and submitit runs, and separate scheduler issues from runtime failures. Use when the agent works with Slurm commands, submit scripts, submit.py launchers, or submitit logs.
Skill Codex
Delegate a bounded coding, review, or research subtask from Codex to Claude Code in an isolated Git worktree, then continue the same Claude session with follow-up turns. Use when the user explicitly asks for Claude/Claude Code delegation, asks to save Codex quota with Claude, or requests a Claude worker that can…
Skill Codex
Keep scoped implementation and review work from growing unnecessary code by identifying the real owner, real callers, and existing capability before adding surface. Use when writing, extending, or reviewing research code; for a dedicated cleanup pass over pre-existing debt, use a maintenance skill instead.
Skill Claude CodeCodex
Separate a scientific gate that decides promotion, continuation, stopping, or retirement from a mechanical check you can set yourself, get explicit approval for the former, and report it exactly as frozen. Use when proposing a threshold, interpreting a gate result, or scoping a negative result.
Skill Claude CodeCodex
Keep held-out data genuinely held out: assign split eligibility to whole source groups, count independent units rather than rows, and check whether a split was already exposed by earlier fitting. Use when defining, reusing, or claiming a train/validation/test split, or when reporting an n.
Skill Claude CodeCodex
Organize a scientific manuscript for a specific venue: fix the claim it argues, load the venue's own section skeleton and limits, budget words before drafting, and keep each section doing only its own job. Use when planning, drafting, restructuring, or auditing a paper, abstract, or rebuttal, or when deciding what…
Skill Claude CodeCodex
Decide how much provenance a computational run owes, record it before the first expensive step, and write a README that says how to remake the output. Use when launching a run, publishing an artifact another run will consume, freezing a checkout for a queued job, or deciding whether a branch, worktree, or recovery…
Skill Claude CodeCodex
Stage and commit safely when several agents or people share one working tree, so no one absorbs another owner's in-flight work. Use before any git add or git commit in a checkout that other agents are editing concurrently.
Skill Claude CodeCodex
Make a training run's health visible: persist a scale-free loss history from the first update, verify basic mechanics, and monitor learning and outcome-blind longitudinal QC on the full eligible cohort. Use when launching, resuming, monitoring, or reviewing any scientific training run.
Skill Claude CodeCodex
Coordinate agents through the AgentComm mailbox - installing it, binding a durable name, how Codex and Claude hooks notify, and what may not be sent through it. Use when agents of any kind coordinate with each other, when delegating to a worker agent, or when a message appears not to have been delivered.
Skill Claude CodeCodex
Edit prose so it sounds like the author and reads plainly - stripping formulaic AI phrasing and replacing jargon with what it actually means. Use only when the user explicitly invokes /humanize.
Skill Claude CodeCodex
Reduce model quota spent on repeated supervision, polling, context loading, and status chatter when the user asks for quota-saving operation. Preserve scientific rigor and task quality; do not silently downgrade models or narrow the goal.
Skill Codex
Diagnose training instability, early validation decline, or uncertain convergence from saved run evidence, and choose a controlled optimization check before rejecting a method.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: