logo.png
// AI Best Practices

Claude Code AI Virtual Staff

A virtual team of 13 agents (product, tech, design, QA, security, growth…) you can plug into Claude Code or OpenAI Codex. An alternative to Superpowers for those who want to keep control over every skill and arbitrate inter-agent friction themselves. Article in 'good to know' mode — I don't use this tool daily, full credit goes to tiber76.

Tool designed by tiber76· 2026-05-10

What it is

A starter kit that installs a multi-agent virtual team into a project. Once `npm run setup` runs, you get 13 specialised agents and 16 orchestration skills, accessible via slash commands in Claude Code or dollar commands in OpenAI Codex.

  • 7 tech agents: po-metier, full-stack-lead, designer-uxui, qa, cso, data-engineer, ai-llm-engineer
  • 6 growth agents: growth-lead, sales-b2b, customer-success, copywriter-brand, content-seo, marketing-analytics
  • 16 skills: setup-project, call-tech-lead, call-growth-lead, redige-us, fullstack-lead-tech, investigate-bug, review-pr, qa-flow, ship-pr, security-audit, redige-brief, ship-landing, audit-funnel, brief-demo, retro-campagne, retro

Each run produces a `TRANSCRIPT.md`: a persistent trace of decisions, trade-offs and deliverables. That's what sets it apart from a simple stack of prompts: traceability is built in.

How it works

An orchestrator receives the user request. It picks the relevant agents based on the topic, gathers an independent opinion from each, arbitrates inter-agent friction (an optional "round 2" if a real tension emerges), implements the deliverables, verifies, and documents.

The mobilisation level is explicit via `--depth=lean|standard|full`:

  • lean — small feature, understood bug, localised refactor. Fewer agents, round 2 rare.
  • standard — normal product feature. Recommended default.
  • full — auth, payment, PII, data model, AI, pricing, major GTM. Quality first, broad mobilisation.

How to install it

Clone the repo, run the interactive setup from this repo, then open the target project:

git clone https://github.com/tiber76/ClaudeCode_AI_Virtual_Staff.git
cd ClaudeCode_AI_Virtual_Staff
npm run setup   # 3 questions: target project, tool (claude/codex/both), conflicts (backup/abort/overwrite)

The script copies the right files, renames the templates, updates `.gitignore`, and generates two reports: `virtual-staff-install-report.md` and `virtual-staff-ai-setup.md`. Once in the target project:

Claude Code  : /setup-project --ai
OpenAI Codex : $setup-project --ai

The AI inspects the repo, infers the stack, the commands, the business entities, the risks, proposes a synthesis, and only asks blocking questions. Recommended first run in `--depth=lean` to calibrate.

When to use it instead of Superpowers

Superpowers (Jesse Vincent) and CC_AI_Virtual_Staff (tiber76) address a similar need — reusing AI patterns — but with two different philosophies.

  • Superpowers — monolithic skills executed by Claude Code itself, focus on daily dev, native Anthropic infra. Lighter for solo dev, ideal if you want "plug & play".
  • CC_AI_Virtual_Staff — multi-agent virtual team, explicit orchestration, traceability via TRANSCRIPT.md, Claude ↔ Codex portability without rewriting skills. Better suited if you want to map agents, retain control over inter-role arbitration, and be able to switch from one LLM provider to another.

Personally I don't use CC_AI_Virtual_Staff daily — Superpowers covers my solo-dev needs. But the multi-agent approach with built-in retro is interesting for larger teams, or for devs who want to understand/modify each skill rather than consume an opaque infra.

Sources & credits

The repo is fully signed by tiber76 — already known on this site for `monitor-ccu` (the Claude Code cost monitoring tool). The multi-agent orchestration patterns and the "optional round 2" mechanic are their contribution. This article is a README read, not a usage report.

Sources & credits