You open a notebook, reach for a data skill, and its session cost becomes part of the choice. The catalogue contains 7,868 skills for this work. The right one depends on its instructions and when your agent loads them.
What loads before you open the notebook
A skill is a folder of instructions the model opens when it decides it needs them. The context is everything the model can see while it answers you. Before a skill is relevant, the model sees its name and description, not the whole folder.
Claude Code accounts for 7,422 of them, Codex for 4,390 and Cursor for 54; a skill can be labelled for more than one. Of the 7,868 on this shelf, 7,608 load their descriptions every session. That is not the full cost, but it shows how much choice your agent carries before you ask about a notebook or model.
The 64-token median for keeping a skill ready
A token is the unit your AI plan bills by, about three quarters of a word. The median skill costs 64 tokens every session. A 64-token preload is about the size of a short note in your prompt box.
The p90 is 151 tokens every session. The largest preload is multi-source-data-integrator, at 685 tokens across 95 lines. Open 20 sessions in a day, and the 64-token median becomes 1,280 tokens before work begins.
What the safety scan can put in front of you
The scan grades 7,402 skills A, 280 B, 185 C or D, and 1 E or F. It finds network calls in 692 skills, root access requests in 111, and remote code downloads and execution in 80. Another 69 use instruction-override phrasing.
The inference-sh-cli is graded C for downloading remote code and making network calls. advisor-orchestrator-worker and understand-knowledge are graded C for recursive force deletion.
The scan prints each finding beside the line that produced it. Read that line in context, then decide whether the behaviour belongs in your project.
Six starting points for data and model work
The original is the first copy of a file we saw; everything matching it afterwards is a copy.
- bootstrap-realtime-eval: starts a new realtime evaluation inside the OpenAI cookbook repository, with its prompts, tools, data and graders. 76 tokens every session, 134 lines, original.
- mle-workflow: turns machine-learning experiments into systems that can be trained, evaluated, deployed, monitored and rolled back. 44 tokens every session, 347 lines, original, 6 copies.
- claude-api: a reference for building on the Claude API and the Anthropic SDK. 294 tokens every session, 581 lines, 11 copies.
- model-integration: guides the addition of a new model or pipeline to Diffusers, including structure and weight conversion. 45 tokens every session, 124 lines, original.
- omni-inference: provides one interface for chat, embeddings, images, speech, moderation, reranking and Responses requests. 47 tokens every session, 1,677 lines, original.
- flow-nexus-neural: trains and deploys neural networks in separate cloud workspaces. 21 tokens every session, 739 lines, original, 5 copies.
When a rule or hook fits your job better
A rule applies guidance only to matching files. A glob is a filename pattern, like *.tsx. Use that narrower mechanism for file-specific notebook or model code.
Use a hook for an event-driven action. A hook is a shell command your agent runs for you when something happens. A skill can explain a script, but it does not react automatically to a file change.
Questions people ask
Which skill should I choose for a full machine-learning system? Choose a workflow covering training, evaluation, deployment and monitoring. One such body reaches 4,645 tokens. Its broader guidance can justify a small preload when the task needs it.
Will a skill load its whole body before I ask for it? Not always. 260 skills wait to be invoked, and the median on-demand body is 1,092 tokens. Check whether the page says it loads every session or waits to be asked for.
Does an A grade mean I can skip reading? No. Of these, 35 have been reviewed and 1 is marked high risk. Treat the grade as a reason to inspect the printed evidence, especially when the skill can affect data or model files.
Do it now
Run npx agentmods cost to see what the mods in your project cost per session, then browse data and AI skills for the job you need.