세션 전반에 걸쳐 컨텍스트 관리를위한 Claude Code 기술. letting /compact 대신 모델이 기억하는 것을 결정하고, 짧은 curated handoff 메모를 저장하고 깨끗한 창으로 다음 세션을 시작, 그 노트만 읽으십시오.

이 메서드는 Anthropic이 structured note-taking 에 대한 효과적인 컨텍스트 엔지니어링에 대해 설명합니다. 지속적 노트는 컨텍스트 창 밖에 유지하고 나중에 세션에서 다시 끌어 들였습니다.

왜 /compact?

• /compact 킥은 컨텍스트가 이미 넣을 때, 그 시점에서 모델은 "dulled"이며, 그 자체를 결정합니다.
• 모든 압축은 무언가를 잃습니다 : 사진을 압축하고 세부 사항을 잃고 세션을 압축하고 결정을 잃습니다. 컴팩트 한 후 컴팩트 한 photocopy의 복사입니다.
• Handoff는 deliberately 및 초기 (약 50-60 %의 컨텍스트 사용), 한 선택적 패스. 프로젝트의 일반 마크 다운 파일이므로 검토하고 수정할 수 있습니다.

어떻게 작동합니까?

1. 세션은 무거운, 그래서 당신은 유형 /handoff. 에이전트는 `session-logs/YYYYY-MM-DD-topic에 구조된 메모를 작성합니다.

handoff-skill: /handoff + /pickup for Claude Code

A Claude Code skill for managing context across sessions. Instead of letting /compact decide what the model remembers, you save a short curated handoff note and start the next session with a clean window, reading only that note.

The method matches what Anthropic describes as structured note-taking in Effective context engineering for AI agents: persistent notes kept outside the context window and pulled back in later sessions.

Why not /compact?

• /compact kicks in when the context is already stuffed, and at that point the model is “dulled” and decides on its own what matters.
• Every compression loses something: compress a photo and you lose detail, compress a session and you lose decisions. Compact after compact is a photocopy of a photocopy.
• A handoff is written deliberately and early (at roughly 50-60% context usage), in one selective pass. It is a plain markdown file in your project, so you can review and correct it.

How it works

1. The session gets heavy, so you type /handoff. The agent writes a structured note to session-logs/YYYY-MM-DD-topic.md: summary, decisions, state (done / in progress / blocked), artifacts, the next step, and what NOT to do (dead ends).
2. You clear the window (/clear) or open a new terminal.
3. In the fresh session you type /pickup. The agent reads only the newest handoff (not the old conversation), verifies reality (git status / ls) and summarizes: where we are, the next step, what to avoid.

Installation

git clone https://github.com/simplybychris/handoff-skill.git
cd handoff-skill

mkdir -p ~/.claude/skills ~/.claude/commands
cp -R skills/session-handoff ~/.claude/skills/
cp commands/handoff.md commands/pickup.md ~/.claude/commands/

From now on /handoff [optional-topic] and /pickup [optional-topic] work in every project.

When to hand off (rough thresholds)

• Models with smaller windows: around 100k tokens.
• A standard 200k window: around 200k tokens is already the risk zone, aim earlier.
• A 1M window: around 300-400k tokens.
• Universal rule: ~50-60% of the context meter, not 90%. And do not interrupt a working implementation mid-step just because the meter grew; finish the micro-step, then save.

Research (including “Lost in the Middle” and Chroma’s test of 18 models) shows that the longer the context, the more often a model loses information from the middle of the conversation, even on trivial tasks.

Repo structure

handoff-skill/
├── skills/session-handoff/SKILL.md   # the skill (mechanics, file format, rules)
├── commands/handoff.md               # /handoff, save mode
└── commands/pickup.md                # /pickup, resume mode

License

MIT