Shared agent skills and configurations across Grok Bot, Grok Build, Claude Code, and other platforms.
- Python 68.1%
- Go Template 24.8%
- Jinja 4.1%
- Shell 3%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
Grok
99bc815347
Agents were validating Vaultwarden by ad-hoc bw unlock, and bash-sourcing the host dotenv which breaks on unquoted shell metacharacters in sibling keys. Add a catalog skill plus a validate helper that parses the dotenv, API-key logs in, unlocks, syncs, and reports counts/names only. Host URL and env-var names live in configs/_shared/vaultwarden.md so the skill body stays bindings-only. Ignore Python bytecode so helper runs do not stage __pycache__. |
||
| agents | ||
| configs | ||
| design | ||
| docs | ||
| overlays | ||
| skills | ||
| standards | ||
| .gitignore | ||
| README.md | ||
agents
Shared standards, skills, and agent configurations for use across platforms:
- Claude Code
- Grok (project agents; Grok Bot and Grok Build for host wiring)
- Cursor (and similar)
This repo is the source of truth for portable agent behavior. Platform-specific wiring lives under
configs/; reusable multi-step recipes live under skills/.
Layout
| Path | Purpose |
|---|---|
standards/ |
Norms stated once, cited by everything else. Start here. |
agents/ |
Portable role contracts — what an agent is for, with no stack or project detail |
overlays/ |
Stack rules (Godot, Python/Qt). Platform-independent; appended to a role |
skills/ |
Portable skill packs (SKILL.md + helpers), with a bindings table for host paths |
configs/claude-code/ |
.claude/ frontmatter dialect, permission grant sets |
configs/grok/ |
.grok/agents/ frontmatter dialect, two-role mapping |
configs/grok-bot/, configs/grok-build/ |
Bot seat mapping / Build session wiring onto agents+overlays |
configs/cursor/ |
Cursor rules and conventions |
configs/_shared/ |
Org forge URLs, token env names, label taxonomy |
docs/ |
How to consume and publish shared agents |
design/ |
Visual language packs (homelab apps, etc.); Design owns, Code Apps/Web cite |
How the pieces combine
An agent file in a consuming repository is four parts: host frontmatter, a role contract, a stack
overlay, and a short project block naming that repository's documents and known state. See
docs/composition.md.
Rules
- No secrets, tokens, or passwords in this repo. Point at secret stores by name only.
- Keep skills and roles generic; put host-specific paths and IDs in
configs/<platform>/, and project-specific state in the consuming repository. - Cite a standard, never transcribe it. A transcribed norm is a copy that will drift — the label taxonomy reached five hand-written copies before this repo existed.
- Prefer small, composable pieces over one mega-prompt.
- Document the consumer: which file each platform should load.
Quick start
- Clone this repo (or add it as a submodule / worktree) where your agent host can read it.
- Read
configs/<platform>/README.mdfor that host's frontmatter dialect and load paths. - Compose each agent from
agents/+overlays/+ your project block, perdocs/composition.md. - Copy or symlink the skills you need from
skills/(includingclean-merged-branches/), and fill in their bindings tables. - Cite
standards/secrets-and-credentials.md; never commit tokens. Claude grant sets: apply by hand fromconfigs/claude-code/settings-permissions.md. Org forge facts:configs/_shared/.