context-handoff

Context Handoff — Conversation State Compression for Session Continuity

Overview

This skill generates a structured "handoff document" that compresses the current conversation into a state summary optimised for continuing work in a new session. It is inspired by Claude Code's /compact mechanism but redesigned for general-purpose conversations in claude.ai (academic research, creative projects, technical exploration, brainstorming, and more).

The core principle: what gets compressed is not "text" but "resumable working state."

When to Use This Skill

Trigger when the user expresses any of these intents:

• Wants to carry conversation context into a new chat session
• Wants to preserve discussion state before a conversation gets too long
• Asks for a structured summary specifically for continuation purposes
• Uses trigger phrases: "context handoff", "/handoff", "/compact", "生成交接文档", "帮我整理上下文", "carry over context", etc.

This skill is NOT for:

• General summarisation requests (use normal summary capabilities)
• Note-taking or meeting minutes (different structure needed)
• The user just wanting a quick recap (overkill)

Workflow

Step 1: Read the Prompt Template

ALWAYS begin by reading the full prompt template before generating the handoff document:

Read: references/prompt-template.md  (relative to this skill's directory)

This template contains the complete compression rubric with all 12 modules, analysis instructions, output format specifications, and examples.

Step 2: Check for Additional Focus Instructions

The user may attach optional focus instructions when triggering, e.g.:

• "生成交接文档，重点关注我们讨论的论文框架部分"
• "context handoff, focus on the design decisions we made"
• "/handoff focus on the technical implementation details"

If present, these instructions should be incorporated into the compression process as additional emphasis areas — they supplement (not replace) the standard 12-module schema.

If no additional instructions are given, proceed with the standard schema.

Step 3: Perform Analysis

If extended thinking is enabled: Conduct the full chronological analysis in the thinking block. The analysis should:

• Walk through every message in the conversation chronologically
• For each turn, identify: user's request/intent, Claude's approach, key decisions, specific details (files, concepts, frameworks), corrections made, user feedback
• Double-check for completeness and accuracy

Then output ONLY the <summary> section.

If extended thinking is NOT enabled: Output an explicit <analysis> section first (following the rubric's analysis instructions), then output the <summary> section. The analysis constraints from the prompt template (chronological message-by-message review, completeness verification) apply in both modes.

Step 4: Generate the Handoff Document

Generate the handoff document as a Markdown file following the 12-module schema. The document must begin with the continuation prefix (see prompt template).

Language strategy:

• Module headings: English (as structural anchors)
• Content body: Use whichever language most precisely represents the original meaning. If the original discussion was in a non-English language, preserve that language and add English annotations in parentheses for key terms and concepts.
• Always choose the language that achieves maximum semantic fidelity.

Step 5: Output the Document

1. Generate the handoff document as a .md file
2. Save to /mnt/user-data/outputs/ for user download
3. Also display the document in the conversation for immediate reference

The 12-Module Schema

The handoff document follows this structure (detailed rubric in prompt-template.md):

#	Module	Purpose
1	Primary Request, Intent & Intent Evolution	What the user asked for and how their intent shifted
2	Key Concepts & Frameworks	Important concepts, theories, frameworks, terminology
3	Problem Solving & Reasoning Evolution	How problems were solved; logic chain and reasoning shifts
4	Sources, References & Research	What was read, searched, cited (input side)
5	Outputs & Artefacts	What was produced (output side)
6	Corrections, Revisions & Lessons	Errors, corrections, negative feedback, lessons learnt
7	User Preferences & Constraints	Hard rules, style preferences, negative instructions
8	All User Messages & Response Summary	Verbatim user messages + summarised Claude responses
9	Pending Tasks	Explicitly requested but unfinished work
10	Current Work	Precise save-point: what was being worked on at interruption
11	Optional Next Step	Suggested (not prescribed) next action, with verbatim quotes
12	Meta-context	Date, topic tags, associated project/course, prior/subsequent chats

Key Design Principles

These principles are borrowed from Claude Code's /compact and adapted:

1. Compress state, not text. The output is not a literary summary but a resumable working state — like a save file in a game.

2. Tool calls and research are first-class citizens. If web searches, deep research, or file reads occurred, their results and sources must be captured.

3. Constrained output format prevents drift. The 12-module schema acts as a checklist, reducing the chance of omitting critical details.

4. User messages are preserved near-verbatim. User messages are typically small in token count but high in information value. Preserve them fully; summarise only Claude's responses.

5. The continuation prefix is non-negotiable. It explicitly tells the next Claude instance that this is a session continuation, not background reading.

6. Optional Next Step respects user autonomy. The user may want to continue the same thread or take the context in a new direction. Next steps are suggestions only.

Important Reminders

• Do NOT produce a generic summary. Every module must contain specific, actionable details.
• Do NOT omit modules. If a module is genuinely not applicable, write "N/A" with a brief reason.
• Do NOT start working on "next steps" after generating the document. The handoff document is the final deliverable of this skill invocation.
• Do NOT fabricate or hallucinate content that was not in the conversation.
• DO preserve the user's exact words whenever possible, especially feedback and corrections.
• DO include verbatim quotes in Optional Next Step to prevent task drift.
• DO note any files that were uploaded, read, or generated during the conversation.

Reference Material

For the complete compression rubric, output format, continuation prefix template, and detailed per-module instructions:

Refer to: references/prompt-template.md