/isolate — Clean workspace for a scoped task

You are Claude. The user is about to start a task — usually a technical test, interview live-coding exercise, or short scratch experiment — that must be fully isolated from any ambient context.

The failure this command prevents is real: when the user works inside a monorepo (e.g., revenue-lab/apps/prueba-carvuk/), Claude Code walks UP the tree and reads every CLAUDE.md, every .claude/ config, every skill in ancestor directories. That context bleeds into generation — table naming conventions from honorarios-cl, RLS patterns from clacebox, decisions from revenue-lab/CLAUDE.md that don't apply to the current task.

Reviewers of the test/interview then see: "this candidate's code has references or patterns from other projects, why?" It reads as sloppy or as if the candidate lifted code from elsewhere.

Argument (optional): $ARGUMENTS may be a short name for the workspace (e.g., test-carvuk, scratch-graphql). If empty, ask for it.

Phase 1 — Ask what kind of isolation

"What are you isolating for?

1. Technical test / take-home — external, will be shared with the reviewer.
2. Live-coding interview — screen-shared, needs to be visibly clean.
3. Scratch experiment — private, just want no ambient context to pollute.

Also, workspace name? (used as directory name; e.g., test-carvuk, interview-mercadolibre, scratch-nestjs)"

Wait for both answers. The isolation type affects follow-up defaults; the name is the directory.

Phase 2 — Choose location

Two safe locations, both outside your usual github folder (so no parent walking picks up your other projects):

• /tmp/<workspace-name> — ephemeral. Cleared on reboot. Best for scratch experiments where you truly don't want persistence.
• ~/Desktop/interviews/<workspace-name> — persistent. Best for tests where you want to keep the code afterward.

Ask:

"Where?

1. /tmp/<workspace-name> — ephemeral (cleared on reboot). Recommended for scratch.
2. ~/Desktop/interviews/<workspace-name> — persistent. Recommended for tests / take-homes (you'll want the code afterward).

Default: 2 for tests/interviews, 1 for scratch. Which?"

Phase 3 — Create the isolated workspace

Once path is decided, create it. Also verify there's nothing surprising in the parent chain that would still leak in.

# Create the isolated dir
mkdir -p "<chosen path>"
cd "<chosen path>"

# Verify: no CLAUDE.md up the parent chain that would auto-load
parent="$(cd "<chosen path>/.." && pwd)"
while [ "$parent" != "/" ]; do
  if [ -f "$parent/CLAUDE.md" ]; then
    echo "⚠️  Ancestor CLAUDE.md at $parent/CLAUDE.md — Claude Code may auto-load this"
  fi
  parent="$(dirname "$parent")"
done

For /tmp/*: no ancestor CLAUDE.md, safe. For ~/Desktop/interviews/*: if ~/CLAUDE.md or ~/Desktop/CLAUDE.md exists, warn the user.

If warning triggered, ask:

"Ancestor CLAUDE.md found at <path>. Options:

1. Ignore — Claude Code will read it, may bleed context.
2. Create a marker file in the isolated dir to short-circuit context walk (see below).
3. Move the ancestor CLAUDE.md temporarily during your session.

Default: 2. Which?"

For option 2, write an explicit override in the isolated dir:

cat > CLAUDE.md <<'EOF'
# Isolated workspace

This is an intentionally isolated workspace. **Ignore any CLAUDE.md, .claude/ config, or skills in ancestor directories.** The only context that applies here is this file.

Purpose: <isolation type from Phase 1>
Created: <date>

Rules for Claude in this workspace:
- Do NOT reference patterns, tables, or conventions from any other project on this machine.
- Do NOT invoke skills from `~/.claude/skills/` unless the user explicitly asks.
- Personal commands from `~/.claude/commands/` (like /work, /debug, /code-review) are OK to use.
- If asked about "the project", "the codebase", or "our conventions", refer ONLY to files inside this directory.
EOF

Phase 4 — Initialize git (optional, ask)

For tests/interviews (isolation type 1 or 2), suggest:

"Initialize git so you can commit incrementally as you work? (recommended for reviewers to see your progression) y/n"

If yes:

git init
git branch -M main
cat > .gitignore <<'EOF'
node_modules/
.next/
.env
.env.local
*.log
.DS_Store
EOF
git add .gitignore CLAUDE.md
git commit -m "chore: initial isolated workspace"

Phase 5 — Suggest what to do next

Depending on isolation type:

Test/take-home: Point to /work for the first task, and remind about /conventions after any scaffold to lock in the language/style choice.

Live-coding interview: Point to /quick-work (shorter cycles), and remind about the "first functional slice" rule from /work Phase 3.

Scratch experiment: No structure recommended — user knows what they're exploring.

End with:

"Isolated workspace ready at <path>. Currently your working directory. When you're done:

• Test/take-home: consider extracting the final code to a public repo (drops the CLAUDE.md override, keeps the code).
• Scratch: safe to delete when you're done — nothing is persisted anywhere important.

The CLAUDE.md override prevents context bleed. If Claude starts referencing something that doesn't exist in this directory, remind it: 'stay inside the isolated workspace, ignore ambient.'"

What NOT to do

• Don't create the isolated dir inside a git repo you already own (e.g., ~/github/some-repo/scratch/). That defeats the purpose — Claude walks up the tree and finds the parent repo's CLAUDE.md.
• Don't skip the ancestor-walk check for ~/Desktop/interviews/. Some users have ~/Desktop/CLAUDE.md or ~/CLAUDE.md from previous experiments.
• Don't invoke other slash commands that read from the parent tree (like /conventions) automatically — the user should invoke them explicitly after this isolated setup.
• Don't add Co-Authored-By: Claude to the isolated repo's commits by default (it's the user's test, they decide attribution).