DEV Community

Dan Evans
Dan Evans

Posted on

I built a framework that gives Claude Code 11 specialist agents and structured multi-agent review

The problem

Most AI-assisted coding today is "vibe coding" — you prompt, the AI writes, you accept. No review. No independent evaluation. No captured reasoning.

The code works (maybe), but nobody understands why it was built that way.


What I built

An open-source framework template for Claude Code that enforces a disciplined development loop:

Plan → Build → Review → Learn → Ship

11 Specialist Agents

Agent What It Does
Architecture Consultant Structural alignment, component boundaries
Security Specialist Vulnerabilities, auth patterns, threat modeling
QA Specialist Test coverage, edge cases, reliability
Performance Analyst Latency, scalability, resource efficiency
UX Evaluator Interaction flow, accessibility, platform conventions
Independent Perspective Anti-groupthink, hidden assumptions, unconsidered alternatives
+ 5 more Docs, education, project analysis, lineage tracking, facilitation

16 Slash Commands

  • /plan → Spec-driven feature planning
  • /build_module → Build with mid-build checkpoint reviews
  • /review → Multi-agent specialist code review
  • /deliberate → Structured multi-agent discussion
  • /ship → Full release workflow with quality gates
  • /walkthrough → Guided code explanation
  • /quiz → Comprehension assessment
  • ...and 9 more for discovery, analysis, retrospectives, and knowledge management.

The Principle I Care Most About

The agent that writes code is never the sole reviewer.

Every change gets independent evaluation from specialists who didn't participate in generation. This single rule prevents the "AI echo chamber" where the same model approves its own work.

Education Gates

Code doesn't merge until you understand it:

Walkthrough → Quiz → Explain-back → Merge

The AI doesn't just build for you — it teaches you what it built and why.

Four-Layer Capture Stack

Every decision, trade-off, and discussion is recorded:

  1. Immutable logs — sealed event files per discussion
  2. SQLite index — queryable metrics and relationships
  3. Curated memory — human-approved patterns and lessons
  4. Optional vector — for when the corpus grows large

Who it's for

Developers who want more rigor from AI-assisted development — not less AI, but smarter AI collaboration.

People learning to code who want AI that teaches as it builds — every decision is explained, every trade-off is documented.


Try it

Here are three links to get you started:

  1. Framework Presentation — interactive overview
  2. How-To Guide — getting started
  3. GitHub Repo — clone the template

GitHub logo Diviner-Dojo / agent_framework_template

AI-Native Agentic Development Framework — a Claude Code template with 11 specialist agents, structured multi-agent review, and automated quality gates

Apache 2.0 licensed. Feedback and contributions welcome.

Top comments (0)