client-boundary-guard

client-boundary-guard

Prevent cross-client data leakage in agency or consulting work. Keep client content, code, credentials, and patterns isolated.

Purpose

When working with client material alongside personal projects, ensure no client data leaks into personal repos, Notion, Linear, or skills — and no personal project data leaks into client deliverables.

When to use this

• Working on a client project alongside personal projects
• Extracting reusable patterns from client work
• Importing or exporting data between client and personal workspaces
• Setting up a new client project

When NOT to use this

• Do not use for ordinary single-project work with no cross-client or cross-workspace boundary risk.

Process

1. Client project isolation

Each client project should have:

• Its own GitHub repo (never mixed with personal code)
• Its own Linear workspace or team (never shared issues)
• Its own Notion space or isolated section
• Its own environment variables and credentials
• Its own CLAUDE.md (never referencing other clients)

2. Pattern extraction rules

When you learn something useful from client work that you want to reuse:

• Extract the principle, not the implementation
• Never include client-specific code, data, or naming
• Never reference the client by name in reusable skills
• Frame as: "When building [generic category], [principle]" — not "From the Acme Corp project"

Example:

• Bad: "Use the same onboarding flow we built for ClientX"
• Good: "Multi-step onboarding with progress indicator, email verification before account access"

3. Context switch protocol

When switching between client and personal work:

• Clear Claude Code context (/clear)
• Verify CLAUDE.md loaded is for the correct project
• Never carry conversation context from one client to another

4. Credential isolation

• Each client's API keys, database credentials, and env vars live in that project only
• Never reference one client's credentials from another project
• Never store client credentials in personal password managers alongside personal ones without clear labeling

Output format

When reviewing a boundary concern:

# Boundary check — [what was reviewed]

## Leakage found: [yes/no]
- [Specific instance]: [where it is, what leaked, fix]

## Isolation status
- Code: [isolated/shared — issue if shared]
- Tasks: [isolated/shared]
- Knowledge: [isolated/shared]
- Credentials: [isolated/shared]

## Remediation
- [ ] [Action needed]

Hard rules

• Never reference a client by name in reusable skills or personal project code
• Never share credentials across client boundaries
• Never carry Claude Code context between client sessions — clear first
• When extracting patterns, extract the principle not the implementation
• When in doubt about whether something is client-specific, treat it as client-specific

Connects to

• pattern-learner — for safely extracting reusable patterns from client work
• memory-import-sanitizer — for cleaning imported data
• permission-guardian — client data operations are Tier 3+
• notion-brain — for maintaining isolated client knowledge spaces

Common failure modes

Pattern extraction includes client details — "Use the dropdown component from the Acme project." This leaks the client name and implies copying their code. Extract the principle instead.

Shared Linear workspace — Client issues mixed with personal issues. One client sees another's task names. Use separate workspaces.