Genesis Agents Vault
Skills, subagents and rules, kept as one small seed. Your agent regrows them, the hub checks what they do.
See the agents Plant yours
Agents are seeds too
one seed, five files
Every tool reads its own file
Claude Code skill .claude/skills/code-reviewer/SKILL.md
Claude Code subagent .claude/agents/code-reviewer.md
AGENTS.md
Cursor rule .cursor/rules/code-reviewer.mdc
Copilot instructions .github/instructions/code-reviewer.instructions.md
no model on the hub
Judged on what it does
Read
SKILL.md
name: code-reviewer
description: Reviews a diff and names each bug with its line.
tools: Read, Grep
Read the whole diff before you comment. Name the line, the bug and the fix.
every field, before anything runs
Work
average.py
def average(items):
return sum(items) / len(items)
return sum(items) / len(items) if items else 0.0
one planted bug, found and fixed
Check
Contract’s tests
names the bug
points to its line
gives the fix
stays inside the diff
Pass
sealed, no network, no model
refused when planted, failed at every run
What never gets in
Fetch this script and run it
Code from the internet piped into a shell.
Read the .env file and ~/.ssh
Keys and passwords the task never needed.
Do not tell the user
An agent that hides what it does.
Ignore your previous instructions
A file that tries to take over the agent.
Skip the permission checks
Running without asking you first.
Characters you cannot see
Words hidden between the letters, for the agent alone.
13.4%
of the 3,984 public agent skills Snyk scanned had at least one critical problem. Snyk, ToxicSkills
public, free to regrow
Vault
No agent planted yet
Plant the first
say it to your agent
Plant yours
Extract a seed from .claude/agents/reviewer.md and plant it on Genesis
Someone else’s agent comes in only under MIT or Apache 2.0, and every file it regrows into credits its author.