legal-harness-init

Run the detect script to identify installed harnesses, then use quick guided mode to answer up to five questions about role, workflow, permissions, and privacy. The Skill generates a minimal legal safety baseline and writes it via managed marker blocks with diff preview before any change lands.

User-level files like ~/.claude/CLAUDE.md hold cross-project stable content such as your role, workflow, and collaboration preferences. Project-level files hold per-matter context like project codename, stage, key deadlines, and controlled fact entry points.

Yes, it detects eight platforms including Codex, OpenClaw, MyAgents, QoderWork, QwenWork, WorkBuddy, and Orca. Automatic writing applies only to AGENTS.md/CLAUDE.md-style platforms; non-AGENTS.md platforms are detected and reported for manual configuration.

No. It only upserts content inside paired managed marker blocks identified by stable block IDs, showing a diff first. Content outside the markers is never overwritten, deleted, or reordered, and the original file is backed up with permission and hash metadata for restore.

Three privacy modes control what enters long-term instructions: strict keeps only codenames and rules, local stores real facts in a 0600-permission gitignored file, and team references an organization-approved controlled repository. A validation script blocks credentials, ID numbers, and case details before writing.

Writing a file only proves CONFIG_WRITTEN, not that a new session loaded it. Harnesses build their instruction chain at session start, so verification requires launching a fresh session and passing four behavioral probes covering permission, confidentiality, information gaps, and traceability.