[

[
"# When your $HOME disappears mid-session: building persistent state for a Claude plugin\n\nI've been using Claude Code daily for months and couldn't tell whether I was genuinely improving or just getting faster at the same handful of habits. Anthropic published a study in February that classified 11 observable collaboration behaviors across 9,830 Claude conversations. I wanted to know what mine looked like against that baseline.\n\nSo I built skill-tree: it analyzes your Claude Code or Cowork session history, classifies those same 11 behaviors, assigns one of seven archetype cards (rendered as tarot cards with museum art), and picks a behavior you haven't touched as a growth quest for your next session. The interesting engineering problem turned out to be state persistence.\n\n*The dual-path problem\n\nClaude Code has a stable home directory, so persisting growth quests across sessions is straightforward: write to ~/.skill-tree/. Cowork doesn't. Its $HOME is ephemeral — it gets wiped between sessions. If I wrote quest state to $HOME in Cowork, every session would start from scratch, which defeats the whole point of a \"what haven't you tried yet\" mechanic.\n\nThe fix was a dual-state-path design. In Cowork, the plugin writes to $CLAUDE_PLUGIN_ROOT/.user-state/ instead of $HOME. $CLAUDE_PLUGIN_ROOT survives session boundaries in Cowork where $HOME doesn't. Claude Code takes the ~/.skill-tree/ path. The SessionStart hook checks which environment it's running in and routes accordingly.\n\n

bash\n# Claude Code path (stable $HOME)\n~/.skill-tree/quest-state.json\n\n# Cowork path ($HOME is ephemeral, CLAUDE_PLUGIN_ROOT persists)\n$CLAUDE_PLUGIN_ROOT/.user-state/quest-state.json\n

\n\nThis meant the behavior classification and archetype card logic could stay identical between environments — only the state write path differs.\n\nThe rest of the pipeline*\n\nThe 7-step orchestration runs in 30–60 seconds: find session files, extract user messages, send to a remote classifier (Claude Haiku on Fly.io), assign archetype, synthesize narrative, render the card, return a stable URL. Each analysis gets its own hosted URL — live example at skill-tree-ai.fly.dev/fixture/illuminator.\n\nThe behavior taxonomy comes from Dakan & Feller's 4D AI Fluency Framework (Description, Discernment, Delegation — the fourth axis, Diligence, isn't observable in chat logs). The population baseline is Anthropic's February 2026 study of 9,830 conversations.\n\nInstall in Claude Code:\n

bash\nclaude plugin marketplace add robertnowell/ai-fluency-skill-cards\nclaude plugin install skill-tree-ai@ai-fluency-skill-cards\n

\n\nAlso available as an MCP server (npm install skill-tree-ai) for Cursor, VS Code, and Windsurf.\n\nhttps://github.com/robertnowell/ai-fluency-skill-cards",