agent-twin

agent-twin

English | 繁體中文

A Claude Code plugin that learns who you are from your own conversations, then teaches every future session to talk to you — not a generic user.

agent-twin reads back the AI conversations you've already had, looks at them through four independent analytical lenses, and distills what it finds into a short brief. From then on, whenever you load that brief at the start of a session, Claude already knows the rhythm of how you think, the kinds of mistakes you don't want it to make, and the topics where you've changed your mind before.

   CAPTURE           ANALYZE              COMPRESS          LOAD
  ┌────────┐      ┌───────────┐         ┌──────────┐     ┌────────┐
  │ Convo  │ ───▶ │ 4-frame   │ ───────▶│ behavior │ ──▶ │ Every  │
  │  logs  │      │ pipeline  │         │  _brief  │     │session │
  └────────┘      └───────────┘         └──────────┘     └────────┘
/save_session    /run_pipeline                          /load_persona
/extract_gemini   ~35 minutes            ≤80 lines
/counselor        10 subagents

---

What you get

You give it a few of your real conversations. It gives you back:

• A persona brief — a short, action-oriented document the assistant reads at the start of each session so it stops explaining things you already know, stops missing your emotional register, and stops recommending paths that contradict how you actually decide.
• Four detailed profile artefacts sitting underneath the brief — value system, cognitive patterns, knowledge graph, behavioural patterns — that you can read for yourself if you want to see why the assistant is behaving the way it is.
• A repeatable pipeline you re-run whenever you've accumulated enough new conversation data. The profile gets sharper over time; you don't have to start over.

The point isn't to mimic you. It's to give Claude enough context to drop the warm-up questions every session and meet you where you actually are.

---

Visualizing the persona

Phase 3 and Phase 4 don't just produce text — they produce two folder-shaped artefacts deliberately authored for Obsidian's graph view. The knowledge graph emits typed concept, emotion, person, and event nodes connected by seven distinct edge types (tension, cause, derives, contradicts, reinforces, weakens, stands_for); the behavioural model emits one file per pattern (BP-001, BP-002, …) with cross-links between related patterns. Every internal reference is a wiki-link, so opening the output folder as a vault gives you a navigable, clustering map of how concepts and behaviours sit relative to each other in your actual thinking.

[Knowledge graph visualization will be added]

[Behavioral model visualization will be added]

To open: point Obsidian at $HOME/.claude/agent-twin/personalized/results/profile/ as a vault. Open the graph view (Ctrl/Cmd+G). The wiki-link edges between nodes form a typed concept graph; the clusters that emerge reveal how your concepts are grouped — which themes pull on each other, which behaviours share triggers, where your value system tightens around a single anchor. This is the most direct way to see the profile rather than just read its summary; the brief is the operational interface, the graph view is where the structure becomes visible.

---

Get started

You don't need to know git. The plugin installs straight from inside Claude Code.

/plugin marketplace add ctongh/agent-twin
/plugin install agent-twin@ctongh-plugins

Then, in any session:

/counselor

/counselor walks you through a guided conversation that produces enough data for the pipeline. When it finishes, it tells you to run /run_pipeline. Around thirty-five minutes later you'll have your first brief. Open a fresh session and run /load_persona — that's it.

If you'd rather feed it a conversation you've already had, use /save_session (snapshots the current Claude Code session) or /extract_gemini (imports a Gemini share-link conversation) instead of /counselor.

Troubleshooting: install fails with an SSH error

The marketplace fetches plugins by cloning their git repositories. By default it uses SSH, which only works if your machine has an SSH key registered with GitHub. If install bails out with a message about authentication, permissions, or [email protected], you have two options.

Option 1 — set up an SSH key on GitHub. This is the long-term fix and the option to pick if you publish your own plugins. GitHub's official walkthrough lives here: https://docs.github.com/en/authentication/connecting-to-github-with-ssh. Once your key is on your account, retry the install.

Option 2 — tell git to use HTTPS instead, just for fetches. If you only consume plugins and never push code to GitHub, you can sidestep SSH entirely with one line of configuration: