reunion

Overview of reunion skill

What reunion does

The reunion skill helps you create a local memorial chat agent based on a deceased loved one’s memories, language habits, and personality traces. In practice, reunion is not a general chatbot prompt pack: it guides intake, material import, memory analysis, persona building, preview, and then ongoing conversation through a dedicated CLI or MCP server.

Who should install reunion

This reunion skill is best for users who want a structured, local-first workflow for digital remembrance instead of improvising everything with ordinary prompts. It fits people who have chat logs, diary text, photos, or oral recollections and want a repeatable process for Agent Orchestration, not just a one-off emotional conversation.

Why users choose reunion over ad hoc prompting

The main differentiator is its two-track model: Memory for shared events and Persona for speech style, values, and behavioral boundaries. The repository also includes progressive recall, RAG-style retrieval, WeChat parsing support, and a safety guard layer. That makes reunion more install-worthy than a plain “act like my relative” prompt if you care about consistency and local data handling.

How to Use reunion skill

Install reunion in the right context

The repository is designed for local use with Claude Code-style skills plus Python components. The README shows cloning into ~/.claude/skills/reunion on macOS/Linux or %USERPROFILE%\.claude\skills\reunion on Windows. Python dependencies are listed in requirements.txt, including mcp, chromadb, sentence-transformers, rich, and typer. If you want the MCP path, inspect mcp_server.py; for simpler local testing, start with cli.py.

Know the minimum inputs reunion needs

A workable reunion install does not require perfect archives. The core flow asks for:

• a name or form of address
• one-line basic facts: age, occupation, region, time since passing
• one-line personality impression
• optional source material: chat logs, diary text, photos, or spoken recollections

Better inputs produce much better output. Strong source material includes repeated phrases, habits, family routines, values, and concrete shared events. Weak input is abstract praise like “kind and hardworking” without examples.

Turn a rough goal into a strong reunion prompt

If you invoke reunion through an agent, do not say only “help me recreate my grandmother.” Give the workflow enough structure to follow its prompts and builders. A stronger reunion usage request looks like:

• “Use reunion to create a memorial agent for my grandmother.”
• “Display name: Grandma Li; relationship: grandmother.”
• “Basic info: 82, retired teacher, Shandong, passed 3 years ago.”
• “Speech habits: often said ‘eat well first’; frugal, caring, slightly nagging.”
• “Materials: chat.txt, notes.md, and 3 oral memories.”
• “Please analyze both shared memory and persona, then show me a preview before generating files.”

That aligns with prompts/intake.md, memory_analyzer.md, persona_analyzer.md, and the preview-confirm step in SKILL.md.

Read these files first and follow this workflow

For a fast install decision, read in this order:

1. SKILL.md for the intended end-to-end flow
2. README.md for setup and command expectations
3. cli.py for local interaction model
4. mcp_server.py or mcp_server_simple.py for Agent Orchestration use
5. core/chat_engine.py, core/memory_store.py, and core/safety_guard.py for runtime behavior
6. prompts/ for the actual quality bar

Suggested reunion guide workflow:

1. test creation with minimal text input
2. inspect generated memory/persona summaries
3. add richer source material
4. only then expose the chat interface to real users

reunion skill FAQ

Is reunion for Agent Orchestration or just for manual chat?

Both, but the repository clearly supports orchestration better than a static prompt file. The MCP server exposes structured tool-style operations, while the CLI gives a lower-friction local path. If you need a reusable memorial-agent workflow inside a broader agent system, reunion for Agent Orchestration is a credible fit.

Is reunion beginner-friendly to install?

Moderately. The user flow is simple, but the stack is not completely plug-and-play. You may need Python environment setup, dependency installation, and comfort reading cli.py or the MCP server files. Non-technical users can still benefit if a developer installs reunion and wraps the workflow for them.

When should I not use reunion?

Do not choose reunion if you want a generic companionship bot, a cloud-hosted SaaS, or zero emotional risk. This skill is specifically about recreating a memorial presence from personal materials. It is also a poor fit if you have no source material and do not want to provide oral memories, because output quality will become generic quickly.

How is reunion different from a normal character prompt?

A normal prompt can imitate tone, but reunion usage is stronger when you need memory retrieval, incremental correction, local storage, and safety checks. The repo separates persona extraction from memory extraction, which usually yields more stable conversations than a single long roleplay prompt.

How to Improve reunion skill

Feed reunion evidence, not adjectives

The biggest quality gain comes from replacing vague descriptors with examples. Instead of “he was loving,” provide:

• common phrases he used
• how he showed concern
• recurring routines
• a specific story with place, people, and outcome

This gives reunion better persona constraints and more useful retrieval anchors than generic emotional labels.

Watch for the main failure modes

Common reunion problems are predictable:

• overly idealized voice because source material is too sparse
• repetitive answers because memory entries are thin or duplicated
• inaccurate intimacy because relationship details were not captured
• emotionally unsafe replies if the model is pushed toward literal resurrection rather than memorial simulation

Check whether the issue comes from missing data, weak prompts, or an unrealistic expectation of factual recovery.

Iterate after the first output

Treat the first generated version as a draft. Use the preview step to correct speech style, values, and memory emphasis before relying on chat. Then add missed materials and rerun the builders. The presence of prompts/merger.md and prompts/correction_handler.md suggests the skill is meant to evolve through additions and corrections, not one perfect first pass.

Improve deployment quality with the right checks

Before wider use, test reunion with a few targeted prompts:

• “How would you ask if I ate dinner?”
• “What family habit would you remind me about?”
• “Tell me a small everyday memory, not a dramatic life summary.”

These checks reveal whether the reunion skill captured ordinary voice, memory specificity, and boundaries. If it fails, revise inputs before changing code. If the flow is solid but runtime behavior is not, inspect core/chat_engine.py, core/context_memory.py, and core/safety_guard.py next.