Skip to content
Star

Context management at scale

Keep 20+ artifacts installed without overwhelming your AI tools. Lazy mode and the skill router give you scale without noise.

The problem

Your team has been productive with grekt. You've published internal artifacts for code review, testing patterns, deployment procedures, database migrations, API design, and more. But now:

  • Your AI tool loads 20+ agent files on every prompt
  • The context window fills up with instructions for things you rarely use
  • The tool loses focus because it's trying to follow too many rules at once
  • Performance degrades as context grows

More artifacts should make your AI smarter, not slower. The problem isn't having too many artifacts, it's loading all of them all the time.

The solution

1. Identify your core vs occasional artifacts

Some artifacts are needed on every interaction (project rules, coding standards). Others are used once a week (database migration agent, release checklist).

2. Set sync modes

Use --core for always-loaded artifacts. Everything else defaults to lazy mode, which stays indexed but out of context. Every grekt add auto-syncs:

bash
# These load on every interaction (core mode)
grekt add @your-org/project-rules --core
grekt add @your-org/coding-standards --core

# These stay indexed but out of context until needed (lazy, the default)
grekt add @your-org/db-migration-agent
grekt add @your-org/release-checklist
grekt add @your-org/api-design-guide

Core artifacts are copied directly into your AI tool's context directories. Lazy artifacts appear only in .grekt/index for discoverability.

3. Use lazy artifacts on demand

When you need a lazy artifact, invoke it through the skill router:

/grekt db-migration-agent

The AI tool reads the index, finds the artifact, and loads it on demand. Once the task is done, it drops out of context.

Result

With sync modes configured:

your-project/
├── grekt.yaml              # Committed - mode: core or lazy per artifact
├── grekt.lock              # Committed
├── .grekt/                 # Gitignored (downloaded artifacts, index)
├── .claude/                # Only core artifacts synced here
└── src/

Your AI tool's context stays clean with only 2 core artifacts loaded. The other 3 are one command away when you need them.