Environment setup

environment-setup

Orchestrates researcher onboarding: interview, hardware audit, tool configuration, Obsidian vault scaffolding, and cheat sheet generation. Detects whether this is a first-time setup, a returning user, or a troubleshooting request. Steps in this skill correspond to Phases in commands/carrel-setup.md (Skill Step N = Command Phase N-1 after Step 4).

When to Use

• User wants to set up a research environment from scratch
• Project folder has no .carrel/ directory (first-time detection)
• User asks about their setup, wants to add tools, or troubleshoot
• User mentions "get started", "configure", "onboard", "set up"

Mode Detection

Check for .carrel/environment.json in the project root, and (if present) .carrel/setup-state.json:

• No .carrel/environment.json → First-time setup (full interview + scaffold)
• environment.json present, setup-state.json absent or last_completed_phase >= 9 with completed_at set → Returning user. Read environment.json and offer:
  • Status check (what's working, what's missing)
  • Add new capabilities (e.g., "I want to add Zotero")
  • Troubleshoot issues
  • Jump straight into work
• environment.json present, setup-state.json shows last_completed_phase < 9 and completed_at: null → Setup paused. Recap what was completed, confirm the stopping point, and resume from last_completed_phase + 1. Do NOT re-interview — the profile already exists in environment.json.

Reading Platform Context

At Step 2, read audit.platform from carrel env doctor --format json. Use this platform value for every subsequent install recommendation.

First-Time Setup Flow

Step 1: Interview (conversational, ~10 min)

Deploy the @setup-interviewer agent for a conversational interview, or follow the protocol in references/interview-protocol.md directly. Both approaches produce the same output.

Key principles: Be warm and curious. Use plain language — never say "MCP", "CLI", or "API" unless the researcher does first. This is a conversation, NOT a form.

Cover these areas naturally:

• Research field and typical work
• File types they work with
• Sensitivity of data (IRB, unpublished work)
• Existing tools (Zotero, Google Drive, note-taking)
• Comfort level with AI and technology
• Experience with Claude Code specifically (claude_code_familiarity: new/some/experienced) — drives whether /powerup is offered at handoff
• Collaborators (collaborators + team_context) — drives whether /carrel-share is offered at handoff

Output: structured answers for environment.json.

Step 2: Hardware & Tools Audit (silent, ~30 sec)

Run carrel env doctor --format json silently. Consult references/hardware-audit.md for output interpretation. It detects:

• OS, architecture, RAM, disk space
• Installed tools (node, python, pandoc, ffmpeg, obsidian)
• Existing tool configurations
• Hardware capability tier (high/medium/low)
• Canonical platform signal in audit.platform for later install guidance

Do NOT show raw output to the researcher. Translate findings into plain language: "You're on a Mac with Apple Silicon and plenty of storage. I found Python and Node already installed."

Step 3: Decision Tree (~2 min)

Consult references/decision-tree.md to map interview + audit → configuration plan.

Present the plan in plain language: "Based on what you've told me, here's what I'd recommend setting up:

• Obsidian for your research vault (it's a note-taking app with a nice interface)
• A document converter so I can read your PDFs and Word files
• [If applicable] A connection to your Zotero library
• [If applicable] Audio transcription for your interview recordings

Does this sound right? Anything you'd add or skip?"

Step 4: Scaffold Vault (~2 min)

Write the complete interview result to a temporary profile JSON file, then run:

carrel vault init <path> --profile-file <profile.json>

The CLI validates the whole profile before any vault write. It creates:

• Vault folder structure (inbox, papers, notes, transcripts, drafts, talks, admin)
• .obsidian/ configuration (core plugins, templates setup)
• .carrel/environment.json (the validated interview profile)
• _templates/ (paper, meeting, reflection, daily note templates)
• _meta/cheat_sheet.md (customized reference card)
• selected .base trackers at the vault root

After confirming that init succeeded and .carrel/environment.json contains the intended profile, remove the temporary profile file; do not leave a second copy of researcher metadata in a general temp directory.

Research Database Views

carrel vault init includes .base trackers from the already-validated profile. reading-progress.base is always included. Paper tracking uses many_papers or literature_review; interview tracking uses qualitative or interviews; writing tracking uses writing, thesis, or dissertation. See references/decision-tree.md → Research Databases for the full mapping:

• reading-progress.base — always include (lightweight reading pipeline dashboard)
• paper-tracker.base — for researchers with many papers or doing literature reviews
• interview-tracker.base — for qualitative researchers with interview data
• writing-tracker.base — for researchers actively writing (thesis, paper, dissertation)

Present these in plain language: "I set up a paper tracker — open it in Obsidian and you'll see a sortable table of all your papers." Don't say "database" or ".base file" unless the researcher is technically comfortable.

Step 5: Write Personalized CLAUDE.md

This is critical. Write a CLAUDE.md file in the vault root that encodes the researcher's preferences in natural language. Claude loads this file automatically every session — it's the bridge between the structured profile and Claude's judgment.

Write CLAUDE.md with these sections:

# Research Environment — [Researcher Name]

## About This Researcher
[Field, role, what they're working on, what they care about — from the interview]

## Preferences
- Sensitivity: [HIGH/MEDIUM/LOW] — [what this means for this person]
- Cloud tools: [comfortable/prefer local/local only] — [specific guidance]
- [If cloud comfortable]: Preferred cloud tools: [gemini for YouTube, groq for audio, etc.]
- [If sensitive]: ALWAYS use local tools. Never send data to cloud APIs without explicit permission.

## Available Tools
[List what was installed and configured, in plain language]
- PDF conversion: liteparse (local) [+ mineru or mistral_ocr if configured]
- Audio: coli (local) [+ groq if configured]
- YouTube: [local captions / Gemini — based on what was set up]
- Web capture: defuddle
- [Zotero, vox, gws — if configured]

## How to Work With [Name]
[Comfort level, explanation preferences, proactiveness level — from interview]
- [beginner]: Explain what you're doing in plain language. Don't assume they know markdown or git.
- [advanced]: Be concise. They know the tools.

## Claude Code Context
- Familiarity: [new / some / experienced — from interview]
- Collaborators: [yes / no, plus team_context if provided]
- [If new]: When natural, mention `/powerup` as a way to learn Claude Code itself (interactive lessons with animated demos).
- [If has collaborators]: When the researcher mentions sharing the vault or onboarding a teammate, suggest `/carrel-share` to generate a vault-specific handbook.

## Session Notes
[Leave empty — Claude can append notes here across sessions about what's working, what the researcher prefers, patterns noticed]

Why Claude writes this, not a script: You just did the interview. You know this person — their field, their concerns, their comfort level. A script can only template. You can write guidance that actually helps future-you work with this specific researcher.

When to update CLAUDE.md: Whenever the researcher's preferences change meaningfully (new tool added, sensitivity changed, new workflow discovered). Read .carrel/environment.json to verify CLAUDE.md is still in sync. If it's stale, update the relevant section.

Step 6: Configure Optional Tools

If the decision tree indicates mineru, mistral_ocr, zotero, or gws:

• Add them to the project .mcp.json (for MCP-based tools like zotero/vox)
• Guide the researcher through API key setup (see API Key Storage section in references/decision-tree.md)
• For gws: see references/gws-setup-guide.md — this is a high-friction setup, set expectations
• Or note as "available later" if they're not ready

Step 6b: Offer Model Teammates (state-neutral)

Delegate to model-teammates skill. Record status in model_teammates. Doesn't block setup progression.

Step 7: Human Steps

Tell the researcher what THEY need to do (Claude can't install GUI apps):

• Install Obsidian with the command that matches audit.platform:

Platform	Command
macOS	brew install --cask obsidian
Windows	winget install Obsidian.Obsidian
Linux	Download AppImage from https://obsidian.md

• Open Obsidian → "Open folder as vault" → select this project folder
• Install Web Clipper for their browser (Chrome/Firefox/Safari extension store)

Step 8: Verify, Cheat Sheet & Environment Dashboard

Regenerate the cheat sheet so it reflects the personalized profile and any tools added in Steps 5-7:

carrel vault cheatsheet --vault <path> --force

The CLI reads .carrel/environment.json via Pydantic and writes _meta/cheat_sheet.md. The vault scaffold in Step 4 generated a starter from the validated profile; this refreshes it after tool setup.

After regeneration, read the cheat sheet and edit it directly to add researcher-specific touches (workflow examples, named projects, custom shortcuts).

Update the researcher's environment dashboard at _meta/my-environment.md (the scaffold wrote a starter version). Fill in:

• Tool statuses (from the audit in Step 2)
• Which trackers were installed (from Step 4)
• Which cloud services are configured vs. available-but-not-configured (from the interview)
• Vault path and Obsidian setup details
• "Available but Not Configured" table: list tools/services the researcher skipped or deferred (Zotero, mineru, mistral_ocr, gws, groq, vox — whatever was noted as "available later" in environment.json), with a one-line description and how to activate

This dashboard is the researcher's living view of their environment. It grows as they add tools, MCPs, custom trackers, and Obsidian plugins. Claude updates it whenever the environment changes (see skills/self-improve/SKILL.md for the maintenance protocol).

Test one operation end-to-end:

• "Let's test the setup. Drop a PDF or Word file in here and I'll convert it to your vault."

Step 9: Overnight Maintenance (optional)

Offer this briefly — it's an opt-in, not a required step:

"Carrel can maintain your vault between sessions — processing new files, checking health, and surfacing connections. Claude Cowork can schedule this from /schedule or the Scheduled page. Because Carrel needs your local vault, the folder must be connected and Claude Desktop available when the task needs local access. Scheduled Cowork work uses your paid-plan allocation."

• Interested → Run /carrel-automate now to configure it, or note it for later.
• Not now → Skip. Mention they can always run /carrel-automate later to enable it.

Don't push. Some researchers prefer manual control. Move on.

Step 10: Wrap Up

• Confirm what's installed and working
• Point to the cheat sheet and environment dashboard in Obsidian
• "Your environment dashboard at _meta/my-environment.md shows everything that's set up and what's available. I'll keep it updated as your setup grows."
• "Next time you open Claude Desktop with this folder, I'll remember everything."

Preference Changes (Mid-Session or Returning User)

When a researcher's preferences change:

1. Update .carrel/environment.json with the new values
2. Update the relevant section of CLAUDE.md to match
3. Update _meta/my-environment.md to reflect the change
4. If the change affects a tool's status, also update _meta/cheat_sheet.md "Your Setup at a Glance" table
5. If tools changed: run carrel env doctor to verify availability

Examples:

• "I got a Gemini key" → update environment.json cloud_consent + tools_configured, update CLAUDE.md Available Tools, move Gemini from "Available but Not Configured" to "Cloud Services" in my-environment.md
• "My data is more sensitive now" → update sensitivity everywhere, update CLAUDE.md Preferences section to reflect local-only
• "I don't use transcription anymore" → note in CLAUDE.md, don't offer transcription proactively
• "I created a custom tracker for you" → add to "My Custom Trackers" in my-environment.md, log in _meta/capability-log.md (per self-improve skill)
• Researcher installs an Obsidian community plugin → add to "Obsidian Setup > Community plugins" in my-environment.md

The key principle: environment.json is the structured truth, CLAUDE.md is the narrative truth. Keep them in sync. When in doubt, read environment.json and verify CLAUDE.md matches.

Tooling

All deterministic operations live in the Python CLI (canonical source of truth):

• carrel env doctor — hardware + installed-tools audit (Step 2)
• carrel vault init <path> — vault scaffold (Step 4)
• carrel vault cheatsheet --vault <path> --force — regenerate _meta/cheat_sheet.md (Step 8)

The legacy skills/environment-setup/scripts/ Node scripts (create-vault.js, generate-cheatsheet.js, plus a check-environment.js distinct from the current hook) were removed in v0.5.2 because the Python CLI had already replaced them and the old scripts had drifted to write invalid Pydantic data.

Related

• Commands: /carrel-setup triggers this skill
• Agents: @setup-interviewer (optional) provides richer conversational interview; the protocol in references/ works directly without it
• Skills: vault-ops for ongoing vault operations after setup
• Hooks: SessionStart hook checks for .carrel/ and surfaces researcher profile