summarize — Claude Code Skill

A Claude Code skill that records session deliverables (files created/modified and cross-session decisions) to a long-term memory file, enabling reliable context restoration across sessions.

Design rationale

Claude Code already has a built-in auto-memory system that automatically captures behavioral patterns — user preferences, feedback corrections, project background, and external references. This skill focuses on what auto-memory does not capture: a chronological log of what was built and what was decided in each session.

What auto-memory captures	What this skill captures
User preferences & habits	File deliverables (paths + summaries)
Feedback & corrections	Cross-session decisions (what + why)
Project background	Pending follow-ups
External resource references	Session date & topic tags

When to use

Run /summarize when the conversation produced:

• File outputs — any files created or meaningfully modified
• Cross-session decisions — architectural choices, direction changes, plan adjustments

Skip for pure Q&A conversations with no file outputs and no decisions that affect future sessions. Claude will detect this and exit gracefully.

Installation

claude skill install https://github.com/Charpup/summarize

Or manually:

mkdir -p ~/.claude/skills/summarize
cp SKILL.md ~/.claude/skills/summarize/

Usage

/summarize

Claude analyzes the conversation, extracts file deliverables and key decisions, and appends a structured entry to your memory file.

Memory file location

Writes to:

~/.claude/projects/<project-id>/memory/MEMORY.md

<project-id> mirrors Claude Code's project path convention (working directory path with separators replaced by -).

Entry format

---
## [YYYY-MM-DD] Task Topic
**话题**：#tag1 #tag2

**文件产出**：
- `path/to/file.md` — one-line description
(omitted when no files were produced)

**决策**：Cross-session decisions made (omitted if none)

**待处理**：Explicit follow-up items (omitted if none)

Auto-archiving

When MEMORY.md exceeds 200 lines:

1. Lines 51+ are archived to memory/archive/YYYY-MM.md
2. A searchable index is maintained at memory/archive-index.md
3. The main file is trimmed to keep context loading fast

License

MIT