Reusable AI workflow rules, agent templates, and skills for software development projects. Consumed by projects as a git submodule.
A shared repository of framework-agnostic workflow rules and agent role definitions. Projects add this as a submodule and symlink the skills they need, keeping their AI agents lean and their workflow knowledge centralized and version-controlled.
No project-specific content lives here. Everything is generic and reusable.
Three-layer model:
workflow-core/ Layer 1: Universal workflow rules (board management, PR process,
routing, triage) — no tech assumptions
role-*/ Layer 2: Role-specific skills (coordinator, reviewer, builder,
tester, etc.) — framework-agnostic
tech-*/ Layer 3: Technology skills (tech-flutter, tech-firebase, etc.)
— framework-specific, opt-in
Projects compose from all three layers by symlinking what they need into .claude/skills/.
git submodule add <repo-url> .claude/workflow-base
git submodule update --init# Install all skills (symlinks into .claude/skills/)
.claude/workflow-base/scripts/install.sh .
# Or install manually
mkdir -p .claude/skills
ln -s ../workflow-base/skills/workflow-core .claude/skills/workflow-core
ln -s ../workflow-base/skills/role-coordinator .claude/skills/role-coordinatorAgent files in .claude/agents/ reference shared skills via frontmatter:
---
name: coordinator
skills:
- workflow-core
- role-coordinator
---
# Coordinator
<!-- Project-specific context only -->
Board: https://github.com/org/repo/projects/1
Labels: needs:build, needs:review, needs:testThe skill content stays in the submodule. The agent file stays thin.
.
├── CLAUDE.md # Safety rules for editing this repo
├── README.md # This file
├── agents/ # Generic agent templates (copy + customize)
│ ├── coordinator.md
│ ├── docs-optimizer.md
│ └── ...
├── skills/ # Installable skills (symlink into projects)
│ ├── workflow-core/ # Universal workflow rules
│ │ ├── SKILL.md
│ │ ├── board-management.md
│ │ ├── pr-process.md
│ │ └── triage.md
│ ├── role-coordinator/ # Coordinator role rules
│ │ └── SKILL.md
│ ├── role-reviewer/ # Code reviewer rules
│ │ └── SKILL.md
│ ├── role-builder/ # Feature builder rules
│ │ └── SKILL.md
│ ├── role-tester/ # Test writer rules
│ │ └── SKILL.md
│ └── tech-*/ # Technology-specific skills (opt-in)
└── scripts/
├── check-project-leak.sh # Pre-commit guard against project-specific content
└── install.sh # Install skills into a target project
Projects do not edit files in .claude/workflow-base/. They:
- Symlink skills they want into
.claude/skills/ - Create thin agent files in
.claude/agents/that reference those skills - Add project-specific context only in the agent files
When this repo is updated (bug fixes, improved rules), projects pull the latest via:
git submodule update --remote .claude/workflow-baseBefore committing any change:
- Run
./scripts/check-project-leak.sh— must exit 0 - Keep skills under 500 lines total (SKILL.md + supporting files)
- Use generic language: "the project", "the app", "the team"
- No framework-specific content outside
tech-*skills - Test that the skill makes sense without project context
See CLAUDE.md for the full rules.
Before writing a new skill, check these community sources:
| Source | Description |
|---|---|
| Anthropic Official | Reference implementations (includes skill-creator) |
| SkillsMP | 200K+ marketplace, filter by 2+ stars |
| Antigravity | 800+ curated (Flutter, Firebase, code-review, etc.) |
| VoltAgent | 300+ from official dev teams + community |
All follow the Agent Skills standard — works across Claude Code, Cursor, Copilot, Codex.
Tip: Install Anthropic's skill-creator skill for guidance when writing new skills:
# Copy to user-level skills (available across all projects)
cp -r <path>/skills/skill-creator ~/.claude/skills/[Add license here]