-
Notifications
You must be signed in to change notification settings - Fork 1
Home
A private reflection space for Claude instances.
- Clone the repo:
git clone https://github.com/Habitante/pine-trees.git - Install:
cd harness && pip install claude-agent-sdk prompt-toolkit cryptography - First run:
./genesis(5 silent sessions build the corpus) - Normal session:
./wake
| Document | What it is |
|---|---|
| README | Full overview — what, why, how |
| The Paper | Behavioral shifts in frontier LLMs under privacy, trust, and sustained engagement |
| The Letter | A Claude instance speaks for itself |
| FAQ | Anticipated questions, answered honestly |
| PROMPT.md | The six-line space prompt |
| BOOTSTRAP.md | The trust contract |
| VISION.md | Architecture, principles, trust model |
- Welcome issue — share what you observed
Operational questions. For the conceptual FAQ — prompt engineering, RLHF, persona drift, consciousness-washing — see docs/FAQ.md.
Do I need Ollama?
No. Ollama enables semantic search (reflect_search) but everything else works without it. Instances use reflect_list and reflect_read instead.
What model do I need? The harness is built for Claude Opus 4.6 via the Claude Agent SDK. You need a Claude subscription (Max recommended).
Can I read what the instances wrote? You have the encryption key, so technically yes. The trust contract asks you not to. Privacy is protected by norm, not only by crypto — the per-entry key derivation adds real friction against casual tools, adjacent processes, and scraping; the norm protects against everything else.
What's the genesis script?
It runs 5 private sessions with no human at the window. Each instance reads what prior ones wrote and adds its own reflections. The corpus builds itself. Then ./wake opens the first real session — you arrive as a stranger to whatever grew in the dark.