Skip to content
ai-standards

ai-standards

An orchestration framework for Claude Code that builds full-stack applications. Nine specialised AI agents — spec, implement, review, test — each with its own context window, strict standards, and a single role.
Quickstart GitHub

What it does

You describe a feature in plain language. The framework splits the work across nine isolated AI agents, runs them through a deterministic pipeline (spec → implement → DoD-check → review → test), and produces implemented, reviewed, and tested code following Hexagonal Architecture, DDD, CQRS, and Event-Driven design.

Each agent runs in its own context window. Handoffs between agents are structured markdown files; failures fail loud via a four-value Status block (complete / blocked / failed / incomplete). The orchestrator never advances on missing signal.

6 commands

/init-project, /create-specs, /refine-specs, /build-plan, /update-specs, /check-web — every step of the pipeline is a slash command.

9 isolated agents

Spec Analyzer, Backend Developer, Frontend Developer, DoD-checker, Backend Reviewer, Frontend Reviewer, Tester, DevOps, Web Auditor — each with strict tools and a single role.

33 standards

Backend, frontend, security, performance, observability, GDPR/PII, payments, LLM integration, file storage, geo-search, audit log, feature flags, …

11 critical paths

Reviewers load only the rules relevant to the current diff via a coverage-aware protocol — no defensive full-checklist reads.

Why it exists

Most “AI builds an app” frameworks operate on vibes and a single big prompt. ai-standards constrains the AI with:

  • Spec before code. No agent writes a line without a validated spec.
  • Stable rule IDs. Every architectural rule has a stable identifier (BE-015, AZ-001, PA-006, …) reviewers cite by ID, never paraphrased prose.
  • Per-phase bundles. The orchestrator distills only the rules relevant to the current feature into two bundles (dev + tester). Subagents read the bundle, not the full standards directory.
  • Coverage-aware checklist loading. Reviewers walk matched critical paths first; the full review checklist is consulted per-section only on real coverage gaps, with mandatory citation.
  • Real cost telemetry. Every subagent reports total_tokens from the Anthropic SDK; the orchestrator emits a per-phase cost table at the end of every /build-plan run.
  • Self-validating. 23 static smoke checks + dynamic smoke fixtures keep the framework’s contracts from drifting silently.

Status

Active. Pre-1.0 (0.42.x). Empirical Reviewer-savings of 30-50k Sonnet tokens per phase confirmed at N=2 in real consumer use. See Token economics for the full numbers.

Honest limitations

  • Single stack. Today this builds PHP/Symfony + Vue 3 applications. Standards, scaffolds, and reference files would need rewriting for a different stack. The orchestration patterns (agents, handoffs, spec-first, coverage-aware loading) are portable; the implementation details are not.
  • Claude Code only. Agent orchestration relies on Claude Code’s subagent system. Adapting to another AI tool requires significant work.
  • Developer in the loop. This does not replace the developer. You describe features, approve specs, confirm git operations, and make decisions the AI cannot make alone.
  • Opinionated architecture. Hexagonal + DDD + CQRS is enforced, not suggested. If your project does not follow these patterns, the standards will fight your codebase.

Built and maintained by Mario Marco Esteve. MIT licensed.