Molyanov AI Dev

English | Русский

AI-First Development Framework

Telegram: @molyanov_blog

A practical, intent-driven methodology for building software with
Claude Code and
Codex. It combines proportional planning, durable Project
Knowledge, focused execution skills, and evidence-gated review without forcing every task through
the same heavyweight pipeline.

User-facing artifacts follow the user's language. Technical documentation, code, prompts, and
skill instructions stay in English so the same project remains portable across sessions and
runtimes.

Two runtimes, one source

Claude files are the editable source of truth. Codex files are generated runtime artifacts.

After editing Claude-side sources, regenerate and review the Codex runtime:

Run conversion after editing sources; project Git hooks also run it before commits and check
for drift before pushes. The converter reports conflicts, validation failures, and managed orphans.
On the next project sync it automatically moves its legacy .codex/skills into .agents/skills;
unknown or manually edited files require conflict resolution.

Quick start

Clone the repository:

For a new or empty setup, copy the runtime you use:

If you already use Claude Code or Codex, do not overwrite your whole configuration. Compare this
repository's CLAUDE.md and AGENTS.md with your files and manually add the instructions you are
missing. Update framework skills and agents selectively, remove only obsolete framework packages,
and keep your personal packages and Codex-owned ~/.codex/skills/.system unchanged.

Global skills physically live in ~/.codex/skills, and agents in ~/.codex/agents. Codex
discovers skills through the ~/.agents/skills → ../.codex/skills symlink; published agents
reference files in ~/.codex/skills. If ~/.agents/skills already…