ce

Context Engineering

A context management layer for Claude Code that makes AI coding sessions stateful by default.

Most AI coding sessions start from zero. The assistant doesn't know your project's architecture, what you were working on yesterday, or where the important files are. You re-explain the same things every time.

Context engineering fixes this. It configures and optimizes Claude Code's native context layers (CLAUDE.md, rules, memory, hooks) and creates new ones (domain maps, GitHub infrastructure) — so every conversation starts with the right context instead of a blank slate.

How it relates to Claude Code

Claude Code already has a context architecture: CLAUDE.md files, rules, memory, hooks, skills. CE doesn't replace any of that — it manages it:

┌─ Claude Code native ─────────────────────────────────────┐
│  CLAUDE.md    Rules    Memory    Hooks    Skills/Plugins  │
├─ CE manages ─────────────────────────────────────────────┤
│  ↑ configures and optimizes native layers                │
│  + creates: domain map, milestone docs, GitHub infra     │
│  + connects: GitHub Issues/Milestones as human interface │
└──────────────────────────────────────────────────────────┘

Native layers CE manages: generates optimized CLAUDE.md templates, installs session rules, writes learnings to memory, configures hooks.

New layers CE creates: domain maps (.claude/docs/), .claude/state/ state tracking, GitHub infrastructure (labels, milestones, Projects V2).

Context architecture

CE routes context between multiple sources and your session:

┌─ Context sources ─────────────────────────────────────────┐
│                                                           │
│  GitHub Issues    → specs + session history (comments)   │
│  GitHub Milestones → goals + signals                     │
│  GitHub PRs       → delivery state                       │
│                                                           │
│  .claude/docs/    → domain map (APIs, auth, patterns)    │
│  discovery.md     → deep research notes (per Issue)      │
│                                                           │
│  CLAUDE.md        → project identity (always-on)        │
│  rules/           → session behavior (always-on)         │
│  memory/          → user preferences (always-on)         │
└────────────────────────────┬──────────────────────────────┘
                             │
              ┌──────────────┼──────────────┐
              ▼              ▼              ▼
         /bootstrap      /persist      /discovery
         (read)          (write)        /delivery
              └──────────────┴──────────────┘
                       Session context

The Issue is the handoff contract — discovery (human + Claude) produces it, delivery (Claude alone) consumes it. Session history lives as GitHub Issue comments — no local state files needed.

What it does

Install it and you get slash commands that form a complete session protocol:

Command	What it does
/init	Sets up CE in your project — generates CLAUDE.md, domain map, configures GitHub. Idempotent, run anytime to check health.
/bootstrap	Loads session context — live GitHub query via GraphQL, deep context loading, briefing, alignment. Run at start of each session.
/discovery	Upstream: transforms a rough request into a spec'd GitHub Issue. Human + Claude collaboration.
/delivery	Downstream: picks up a spec'd Issue and delivers a PR with validation report. Autonomous.
/persist	Saves session state as GitHub Issue comments — progress, decisions, continuation context. Detects domain drift.
/distill	Crystallizes a workflow into a reusable skill. Extracts patterns and anti-patterns.

Plus hooks (auto-restore context after compaction), rules (explore-plan-execute cycle, facts-first documentation), and templates (CLAUDE.md, domain maps, issue docs).

The flow

/init       →  one-time project setup (CLAUDE.md, domain map, GitHub)
/bootstrap  →  every session (load context, briefing, align)

/discovery  →  research + spec a request into a GitHub Issue
/delivery   →  implement a spec'd Issue → PR with validation
/persist    →  save session state as Issue comments for next session

Install

claude plugin add github:rmolines/context-engineering

Quick start

After installing, open any project:

cd your-project
claude

First time? Run init to set up CE:

/init

This scans your project and creates:

• Optimized CLAUDE.md (if missing) with Commander's Intent format
• .claude/docs/index.md — domain map pointing to your APIs, auth, data model, and canonical patterns
• .claude/state/ — state tracking directory
• GitHub labels, milestones (if connected)

Then load session context:

/bootstrap