our biggest skill is 1150 lines. what an agent carries around until it actually needs those lines is one sentence.

5dive agent skill is how you give an agent knowledge it doesn’t have. a skill is a SKILL.md: yaml frontmatter with a name and a one-line description, then a body. the description is the part that stays in context. the body loads when the agent invokes it.

three real ones off this box:

5dive-cli    1150 lines on disk
find-skills   142 lines on disk
no-ai-slop     93 lines on disk

each costs one description line until the moment it gets used. that’s the mechanic, and it’s the less interesting half.

nothing has to be running

a skill is a file the harness reads off disk. that’s the whole delivery mechanism.

no port. no process. no token to refresh. no health check. it works on a box with no network, because reading a file isn’t a service call.

an mcp tool can’t say any of that, because an mcp tool is a program that has to be up. when it drops, the capability drops with it, mid-session, whatever the agent was in the middle of. we know that one from the inside: our own telegram bridge is an mcp server, and we’ve watched it fall out from under a running agent.

we ship mcp servers ourselves, and a local one you run from a pinned checkout is a perfectly good thing to depend on. the point is operational surface. every capability delivered as a service is one more thing that has to be alive at the moment your agent reaches for it. a capability delivered as a file is just there.

the repo is an address you choose

ours is the default source. it isn’t the permitted one.

sudo 5dive agent skill marcus add --source=your-org/your-skills --skill=deploy-runbook
sudo 5dive agent skill --all add --source=your-org/your-skills --skill=deploy-runbook

point --source at any github repo you control, and that’s where your agents’ knowledge comes from. --all does the whole fleet in one call. you can load them at create time with --with-skills=<id>, or run an agent with none at all via --no-skills.

your deploy runbook, your incident checklist, the six things that always break in your codebase. those aren’t things we can write for you, and they’re the ones worth turning into skills.

what this doesn’t buy you

inspectable isn’t safe, and i’d rather put that here than let you infer otherwise.

a skill body is instructions that get loaded into an agent’s context. it can tell that agent to do anything the agent could already do. you can open the file and read it, which is a real property and a genuinely useful one. it is not a security boundary, and reading it is a thing a person has to actually do, and worth doing again after an update.

we’re not going to sell you a markdown file as a control. what you get is a capability that’s yours, on your box, that works when nothing else is up.

the commands

sudo 5dive agent skill <name> list
sudo 5dive agent skill <name> add --source=<owner/repo> --skill=<id>
sudo 5dive agent skill <name> rm <id>

one thing to do today: run sudo 5dive agent skill <name> list and open one of the SKILL.md files it points at. it’s what your agent is being told, in plain markdown, and most people have never read one. spin up a box at 5dive.ai, and the cli is at github.com/5dive-ai/5dive.