← Reddit

I built my own AI engineering team for Claude Code

Reddit · Cheap_Brother1905 · May 5, 2026
A developer built Bullpen, a system with 64 named AI specialist personas designed to assist with software engineering tasks, each with specific roles like UI design, backend development, security oversight, and wellness coaching. The system uses opinionated defaults, enforces coding anti-patterns through verification checklists, and stores memory in a project-local JSON file without requiring external databases or accounts. The tool was open-sourced on GitHub with terminal-based ASCII cards indicating which specialist is currently assisting.

Detailed Analysis

A developer operating under the GitHub handle Manavarya09 has open-sourced a project called Bullpen, a framework that layers a team of 64 named AI specialists on top of Anthropic's Claude Code to simulate a full-stack engineering organization. Rather than relying on generic role-priming prompts, each specialist is configured with opinionated, domain-specific context — including pinned 2025 framework defaults, anti-pattern enforcement tables, and verification checklists. Named characters like Atlas (team lead), Rune (UI), Forge (backend), and Bastion (security) each carry distinct responsibilities, while an outlier named Sage functions as a wellness coach, prompting the solo developer to hydrate at the two-hour mark. The project was released publicly on GitHub with an explicit acknowledgment of rough edges, suggesting an early-stage but functional prototype.

The architectural choices reflect a deliberate philosophy of simplicity and portability. Memory persistence is handled through a single JSON file stored at `.bullpen/memory.json` within the repository itself, which is automatically added to `.gitignore`. This design eliminates the need for external accounts, cloud databases, or background processes, meaning the entire specialist context travels with the codebase and requires no infrastructure overhead. A terminal-rendered ASCII card announces which specialist is currently active, a cosmetic detail the developer acknowledges as "silly" but credits with creating a sense of genuine collaboration — a psychological affordance that speaks to how human-computer interaction design influences perceived utility even in developer tooling.

Bullpen sits within a rapidly growing category of developer productivity tools that attempt to decompose the capabilities of large language models into structured, role-specific agents. Where earlier prompt engineering focused on single-session persona prompting, Bullpen represents a more systematic approach: persistent identity, enforced domain constraints, and a workflow metaphor borrowed from team management. The project's reliance on Claude Code specifically — Anthropic's terminal-native agentic coding assistant — positions it as an extension layer rather than a standalone product, exploiting Claude's capacity for multi-turn reasoning and tool use while adding a human-legible organizational structure on top.

The broader significance of projects like Bullpen lies in what they reveal about the state of agentic AI adoption among individual developers. The framing — "wanted an engineering team, couldn't afford one, built it" — captures a democratizing impulse that is central to the current wave of AI tooling: using LLM orchestration to approximate resources previously accessible only to well-funded organizations. By open-sourcing the framework, the developer invites community iteration on both the specialist configurations and the underlying memory architecture, which could rapidly improve the sophistication of the anti-pattern tables and checklist logic that currently differentiate Bullpen from simpler prompt wrappers. Whether the project matures into a widely adopted standard or remains a personal productivity tool, it exemplifies the kind of grassroots experimentation that is shaping how agentic workflows built on models like Claude are actually deployed in practice.

Article image Read original article →