AI coding tools
Claude Code Operating Card
Choose and operate Claude Code with explicit repository scope, permissions, verification, and recovery.
- Format
- Reference
- Level
- Beginner
- Audience
- Practitioner, Developer, Operator
- Owner
- Project42 Editorial
- Review cadence
- Every 30 days
- Prerequisites
- A version-controlled working copy; An Anthropic-supported authentication path
Use it for a repository-local agent loop
Claude Code is a terminal coding agent that can inspect a repository, edit files, run commands, and use configured tools. Choose it when the task benefits from an interactive repository loop and the operator can review requested actions and resulting diffs.
Capability is not authorization. Read the repository instructions, establish the writable paths, protect secrets, and begin with the narrowest permission posture that can complete the task. Treat MCP servers, hooks, and project instructions as separate trust decisions.
Use an evidence-preserving operating loop
Start from a known branch and worktree state. State the requested outcome, in-scope files, checks, and stop conditions before allowing edits. Review each permission request in context; a familiar command can still target the wrong path.
Keep rollback available with normal version-control checkpoints and backups appropriate to the data. If the agent crosses scope, produces unexplained edits, or cannot run the agreed checks, stop and recover the last verified state before retrying.
Task: [one bounded outcome]
Scope: [repository and allowed paths]
Permissions: [read, write, command, network, and tool limits]
Inputs: [instructions and evidence]
Expected changes: [files or behavior]
Verification: [tests, diff review, and observable result]
Stop conditions: [scope conflict, secret request, unsafe command, failing invariant]
Recovery: [checkpoint, revert plan, and evidence to preserve]Expected evidence and verification
Expected evidence includes the initial worktree state, the approved scope, a reviewed diff, command output for the relevant checks, and a concise list of unresolved risks. Verify that every changed file supports the requested outcome and that no secret, unrelated edit, or generated dependency cache entered the commit.
For recovery, stop the session, preserve useful logs, restore only the affected files from a known checkpoint, and rerun the smallest deterministic check. Do not broaden permissions merely to make a failing task disappear.