cohort

Background advisors for Claude Code — keep the foreground LLM honest, focused, and remembering.

What this is

A small framework of skills + hooks + persistent advisor agents that surround a Claude Code session and:

• Catch drift before harmful actions land (PreToolUse intercept)
• Provide external long-term memory the foreground forgets
• Surface relevant context when foreground is about to need it

Built around two observations:

1. The foreground LLM (your main Claude session) drifts — over-engineers, repeats refused actions, loses scope mid-conversation, only catches it when externally pointed out.
2. Even with 1M context, mid-context recall degrades over long sessions. Specialized advisors with focused contexts hold things the foreground forgets.

Status

Design phase. No code yet. See conversation transcript / future docs/design.md for the architecture being explored.

Repo layout (planned)

cohort/
├── skills/         # User-invoked skills (scope-keeper, etc.)
├── hooks/          # Stop / PreToolUse / UserPromptSubmit hook scripts
├── advisors/       # Persistent advisor agent prompts (memory-keeper, context-curator)
├── docs/           # Architecture, design decisions
└── scripts/        # install.sh, helpers

License

MIT. See LICENSE.