Skip to main content
AEP is a file-based protocol that saves successful AI collaboration patterns as Agent Experience Packs — repo-local JSON files your agent loads at the start of every session. No more repeating constraints, preferences, and workflows from scratch.

Quick Start

Install AEP and save your first pack in minutes.

How It Works

Understand the AEP flow from task to saved pack to reuse.

Commands

Learn the four AEP commands: save, apply, promote, inspect.

Schema Reference

Full field reference for AEP pack schema v1.0-exp.

Why AEP

Every AI session starts from zero. You repeat the same instructions:
  • “Do not redesign the layout”
  • “Reuse the existing CSS”
  • “Keep this practical, not over-engineered”
…even when you already solved this exact task last week. AEP fixes that.
1

Complete a successful task with your agent

Work with your AI agent until you get the outcome you want — with the right constraints, workflow, and checks in place.
2

Save the pattern as an AEP pack

Tell your agent: “Use the AEP skill and save this collaboration.” The agent extracts intent, constraints, preferences, and workflow into a JSON pack stored in your repo.
3

Apply packs on future tasks

Next time, tell your agent: “Use the AEP skill and apply relevant packs before you start.” The agent loads and ranks matching packs — your preferences are already in place before the first line is written.
4

Packs improve over time

Promote strong task patterns to project or user scope. Track usage metrics. Merge overlapping packs. The more you use AEP, the better your agents start every session.

Key features

Repo-local storage

Packs live in .agent/aep/ (or agent-specific directories like .claude/aep/) as plain JSON — no external service, no runtime.

Agent-agnostic

Works with Claude, Codex, Gemini, OpenCode, Cursor, and any agent that can read files and follow instructions.

Smart matching

Packs declare languages, frameworks, paths, and domains. Agents score and rank candidates for every task automatically.

Pack evolution

Track how often packs are applied, when they were last used, and promote proven patterns from task to project to user scope.

Before vs. after

Without AEPWith AEP
Repeat instructions every sessionReuse saved patterns automatically
10–20 correction cycles1–2 iterations
Inconsistent outputsConsistent, aligned results
High token costLower cost over time

Build docs developers (and LLMs) love