A plug-and-play multi-agent system for Claude Code and the Claude Agent SDK. Drop the agents and skills in this repo into any project and instantly get a team of role-based AI collaborators — sales, marketing, developer, LinkedIn growth, content, and analytics — that share context, hand off tasks, and run end-to-end workflows.
Why this exists. Modern teams stitch together half a dozen tools to ship a single feature, post a single launch, or qualify a single lead. This framework collapses that surface area into a single conversational interface where specialized agents collaborate under one orchestrator.
- What's inside
- Quickstart
- The agent roster
- The skill library
- Workflows
- Architecture
- Bring your own credentials
- Extending the framework
- Project layout
- License
| Layer | Purpose | Lives in |
|---|---|---|
| Orchestrator | Routes a user request to the right specialist agent(s) and merges results. | agents/orchestrator.md |
| Role agents | Sub-agents that each own a domain (sales, dev, marketing, …). | agents/*.md |
| Skills | Composable, reusable instruction packs (e.g. write-linkedin-post, qualify-lead). | skills/*/SKILL.md |
| Workflows | End-to-end recipes that chain agents and skills (e.g. launch-day, sprint-review). | workflows/*.md |
| Examples | Copy-pasteable prompts that demonstrate each agent and workflow. | examples/*.md |
.claude/ |
Claude Code's auto-discovery dirs so agents/skills load without manual install. | .claude/agents/, .claude/skills/ |
# As a submodule (recommended — easy upgrades)
git submodule add https://github.com/shyamw-cloud/claude-skill.git .claude-team
ln -s .claude-team/.claude/agents .claude/agents
ln -s .claude-team/.claude/skills .claude/skills
# Or clone directly into .claude/
git clone https://github.com/shyamw-cloud/claude-skill.git .claudeclaude --list-agents # should print: orchestrator, sales-agent, developer-agent, …
claude --list-skills # should print: write-linkedin-post, qualify-lead, …> Use the orchestrator. We launched v2 today.
- Draft a LinkedIn post (founder voice).
- Generate 3 cold-email variants for our top-50 ICP.
- Open a PR adding a /changelog page summarising the new features.
The orchestrator delegates to marketing-agent → linkedin-agent, sales-agent, and developer-agent in parallel, then returns one merged report.
Each agent is a self-contained Markdown file with YAML frontmatter (name, description, allowed tools, model). Claude Code loads them automatically from .claude/agents/.
| Agent | Persona | Best at |
|---|---|---|
orchestrator |
Chief of staff | Decomposing requests, delegating, merging output |
sales-agent |
SDR + AE hybrid | Lead qualification, cold email, discovery questions, proposal drafts |
developer-agent |
Senior full-stack engineer | Code, PRs, code review, refactors, test plans |
devops-agent |
SRE / DLC engineer | CI/CD, IaC, deployments, rollbacks, incident response |
marketing-agent |
Growth marketer | Positioning, campaigns, ad copy, content calendars |
linkedin-agent |
Personal-brand strategist | LinkedIn posts, comment hooks, profile rewrites |
content-agent |
Long-form writer | Blogs, newsletters, case studies, SEO outlines |
analyst-agent |
Data analyst | KPI dashboards, funnel analysis, A/B test readouts |
support-agent |
CS lead | Ticket triage, macro-replies, escalation drafts |
Daily-use lens. The roster maps to the jobs a small startup actually does every day: ship code, talk to customers, post content, watch the numbers.
Skills are task-shaped, not role-shaped. Any agent can invoke any skill. Each skill is a folder under skills/<skill-name>/ containing a SKILL.md (with frontmatter) plus optional templates and references.
Currently shipping:
write-linkedin-post— Hook → story → CTA, founder/operator voice presets.qualify-lead— BANT + MEDDIC scoring against an ICP definition.cold-email-sequence— 4-touch sequence with behavior-based branching.market-research— Competitor scan, pricing teardown, positioning gaps.code-review— Security-first PR review with severity tags.release-notes— Turn a PR list into customer-facing release notes.seo-outline— Keyword cluster → article outline with internal links.funnel-analysis— Diagnose drop-off across an acquisition funnel.content-calendar— 4-week calendar across channels with themes.
Run a skill directly:
> /write-linkedin-post we just shipped agent-to-agent handoff in v2
…or let an agent pick the skill:
> linkedin-agent: write something punchy about our v2 launch
Workflows chain multiple agents and skills into a single command.
| Workflow | What it does |
|---|---|
launch-day |
Marketing + LinkedIn + Sales + Developer collaborate on a product launch |
weekly-growth-review |
Analyst pulls numbers → Marketing proposes experiments → Orchestrator drafts the all-hands update |
inbound-lead |
Sales qualifies → drafts personalized outreach → schedules dev demo |
sprint-review |
Developer summarises PRs → Content writes release notes → LinkedIn posts highlights |
customer-escalation |
Support triages → Developer reproduces → Sales briefs the account |
See workflows/ for the full prompt templates.
┌────────────────────┐
user prompt ───▶│ orchestrator │
└────────┬───────────┘
decompose + │ merge
dispatch │ results
┌──────────┼──────────┐
▼ ▼ ▼
sales-agent dev-agent marketing-agent
│ │ │
▼ ▼ ▼
skills/ skills/ skills/
qualify-lead code-review write-linkedin-post
cold-email release-notes seo-outline
Design principles
- One source of truth per role. Each agent owns its domain prompt; skills are composed in.
- Skills are stateless and reusable. A skill never assumes a calling agent.
- Tool budgets per agent.
developer-agentcan write code;linkedin-agentcannot. Frontmatter enforces this. - Human-in-the-loop by default. Any agent that posts publicly (LinkedIn, email, GitHub) drafts and asks for approval before sending.
Public-facing actions are gated behind credentials you provide via environment variables — the framework never bundles secrets.
| Capability | Env var(s) | Used by |
|---|---|---|
| LinkedIn posting | LINKEDIN_ACCESS_TOKEN |
linkedin-agent |
| Email send | RESEND_API_KEY or SENDGRID_API_KEY |
sales-agent |
| GitHub writes | handled by Claude Code's GitHub MCP | developer-agent |
| Analytics pull | POSTHOG_API_KEY, GA4_PROPERTY_ID |
analyst-agent |
| CRM sync | HUBSPOT_TOKEN or SALESFORCE_TOKEN |
sales-agent |
Copy docs/.env.example to .env and fill in only the integrations you need. Agents whose credentials are missing fall back to draft-only mode.
Add a new agent in three steps:
- Create
agents/<your-agent>.mdwith frontmatter (name,description,tools,model). - Symlink it into
.claude/agents/. - (Optional) Add domain-specific skills under
skills/.
Add a new skill:
- Create
skills/<skill-name>/SKILL.mdwith frontmatter (name,description). - The skill auto-discovers — no registration needed.
See docs/AUTHORING.md for the full spec, including the SKILL.md frontmatter schema.
.
├── README.md ← you are here
├── agents/ ← role-based agent definitions
│ ├── orchestrator.md
│ ├── sales-agent.md
│ ├── developer-agent.md
│ ├── devops-agent.md
│ ├── marketing-agent.md
│ ├── linkedin-agent.md
│ ├── content-agent.md
│ ├── analyst-agent.md
│ └── support-agent.md
├── skills/ ← reusable, task-shaped skill packs
│ ├── write-linkedin-post/
│ ├── qualify-lead/
│ ├── cold-email-sequence/
│ ├── market-research/
│ ├── code-review/
│ ├── release-notes/
│ ├── seo-outline/
│ ├── funnel-analysis/
│ └── content-calendar/
├── workflows/ ← multi-agent recipes
│ ├── launch-day.md
│ ├── weekly-growth-review.md
│ ├── inbound-lead.md
│ ├── sprint-review.md
│ └── customer-escalation.md
├── examples/ ← copy-paste prompts
│ ├── sales-day-in-the-life.md
│ ├── developer-day-in-the-life.md
│ └── marketer-day-in-the-life.md
├── docs/
│ ├── AUTHORING.md
│ ├── ARCHITECTURE.md
│ └── .env.example
└── .claude/ ← Claude Code auto-discovery
├── agents/ ← symlinks/copies of agents/
└── skills/ ← symlinks/copies of skills/
MIT — use it, fork it, ship it.