You stop at /review and /migrate when the task is clear. Slash commands suit work you can ask for directly. If the agent should choose guidance itself, you need a skill instead.
Slash commands fit work you can name
The catalogue has 16,759 commands filed under Claude Code. A familiar command name can still hide a specific purpose.
A command narrows a request before you start coding. A requirements command checks whether your feature is clear. A migration command moves selected issues. A review command examines changes already in progress.
The context is everything the model can see while it answers you. A glob is a filename pattern, like *.tsx. A command using one can work on selected files instead of your entire project.
The original is the first copy of a file we saw; everything matching it afterwards is a copy. That label tells you whether to study the source or compare another implementation.
Type the task, then let the body do the work
The command name and description can appear before you invoke it. The whole body enters the context when you invoke it, so the command stays light until you ask it to work.
What slash commands cost before you type them
Of the 42,815 commands, 31,588 load their names and descriptions every session. The other 11,227 wait until you ask for them.
A token is the unit your AI plan bills by, about three quarters of a word. Among commands that load every session, the median costs 17 tokens and the 90th percentile costs 44.
At 20 sessions a day, the 17-token median becomes 340 tokens across a day. The 17-token median is closer to a short note in your prompt box than a full instruction page.
The largest every-session example is clodex, at 714 tokens. On-demand command bodies have a 540-token median when loaded.
What the command scan puts in front of you
The scan grades 40,897 commands A, 1,059 B, 857 C or D, and 2 E or F. The most common finding is a network call, appearing 1,394 times.
Other findings include 813 reads of agent configuration directories, 462 recursive force deletes, and 187 downloads followed by remote execution. The spec-kit assessment is graded D for instruction-override phrasing and a cloud metadata endpoint.
The remove-worktree command is graded C for recursive force delete. The reset command has the same finding.
Each finding gives you the line that produced it. Use it as a pointer for inspection, not as a final judgment on the command.
Six slash commands for named jobs
The figures after each name show token cost, file length, whether the file is original, and its copy count.
- checklist: turns feature requirements into questions before building. 13 tokens every session, 4,883 when loaded, 380 lines, original, 73 copies.
- review-and-fix: reviews a pull request or staged changes and starts a fix cycle. 0 tokens before use, 1,416 when loaded, 100 lines, original, 0 copies.
- migrate: moves selected GitHub issues between repositories. 0 tokens before use, 1,254 when loaded, 185 lines, original, 0 copies.
- gt-workflow: sets a Graphite workflow for branch and commit operations. 0 tokens before use, 865 when loaded, 103 lines, original, 0 copies.
- explain: researches a codebase and creates a compact guide for the agent. 0 tokens before use, 613 when loaded, 71 lines, original, 0 copies.
- js-to-ts: helps migrate a selected JavaScript file to TypeScript. 0 tokens before use, 97 when loaded, 11 lines, original, 0 copies.
When a skill fits better than a slash command
A skill is a folder of instructions the model opens when it decides it needs them. Choose one when the guidance should apply based on the situation, rather than after you remember a command name.
A command cannot choose itself. It waits for your request, so it cannot notice an unfamiliar codebase and load its own instructions.
A hook is a shell command your agent runs for you when something happens. Choose that shape for an event-driven action, not a named task you start deliberately.
Questions people ask
How much context can one command use? The median body is 540 tokens, and checklist opens 4,883.
What should make me pause before installing one? A root request appears in 271 findings. Read the flagged line and the command’s surrounding instructions before proceeding.
Can a command handle work I have not asked for yet? No. If the model should decide when guidance applies, use a skill. If the action should follow an event, use a hook.
Do it now
Run npx agentmods add <slug> to install one command, then browse the command catalogue.