Agentic Engineering Protocol — AI skill for Claude Code

Agentic Engineering Protocol (AEP)

**A verification-driven operating protocol for AI coding agents.** Explore → Plan → Implement → Verify → Deliver — with skeptical brainstorming loops, gap analysis, adversarial review, closed feedback loops, and a hard verify gate.

AEP does not make your model smarter. It systematically closes the failure modes that waste your time with any coding agent: solving the wrong problem, declaring "done" without evidence, grading its own work, silently drifting out of scope, and forgetting what it learned last session.

Works with **Claude Code** (full plugin: skills + subagents + Stop-hook gate + commands) and with **any tool that reads the open Agent Skills standard or AGENTS.md** — OpenAI Codex, Cursor, Gemini CLI, GitHub Copilot / VS Code, and others.

---

Architecture: the right rule in the right layer

Monolithic instruction files degrade: the longer the always-loaded file, the more the agent ignores. AEP splits the protocol across layers so detail is abundant where it's free and discipline is enforced where it matters:

Layer	Mechanism	Loaded	Carries
Always-on core	AGENTS.md (+ thin CLAUDE.md adapter)	Every session, in full	Operating stance, non-negotiables, protocol summary, project memory (§P)
Playbooks	9 skills (aep:*)	On demand (name+description always visible; body loads when invoked/matched)	Deep procedural detail: 5 phases + standards + triangulated research + orchestration
Fresh-context review	4 subagents	On delegation, isolated context	Adversarial review, gap audit, security & performance audits — the author never grades its own work
Hard enforcement	Stop hook (verify-gate.sh)	Deterministic, outside the model	Blocks "task complete" while the project's check fails
Bootstrap	/aep:init command	Manual	Installs the core + gate into any repository, populates project facts

Quick start — Claude Code

# 1. Add the marketplace