This repository contains agent skills and scripts. Treat them as executable workflow assets, not just text.

Scripts should be:

• non-interactive by default;
• clear about inputs and outputs;
• safe by default;
• explicit about file writes;
• explicit about network access;
• explicit about destructive operations;
• runnable with --help.

Destructive operations require explicit user confirmation and should support dry-run when practical.

Do not add network access or new dependencies casually.

Before adding either, document:

network_or_dependency_change:
  reason: ""
  affected_skill: ""
  data_sent_or_fetched: ""
  dependency_or_endpoint: ""
  security_risk: []
  fallback: ""

Skills, references, eval fixtures, and project memory must not store:

• tokens;
• API keys;
• passwords;
• private key paths;
• private machine paths;
• personal identifiers not needed for the task;
• hidden prompt-injection payloads outside isolated test fixtures.

Stateful skills must follow docs/DBX_STATEFUL_SKILLS.md.

Any state write needs:

• owner;
• lifetime;
• update policy;
• stale policy;
• privacy boundary;
• rollback path;
• approval rule for external writes.

Project memory is editable evidence, not eternal truth.

Commands, hooks, status lines, AGENTS.md, CLAUDE.md, llms.txt, planning files, and project memory should be reviewed as control mechanisms.

Hooks require extra caution because they may run without the user explicitly asking at that moment.

A hook is acceptable only when:

• the failure is concrete and recurring;
• the hook is less risky than relying on model memory;
• the user can understand and disable it;
• side effects are documented.

Before installing third-party skills:

1. Read SKILL.md.
2. Inspect scripts/.
3. Check for network access, dependency installs, credential access, and destructive commands.
4. Prefer installing only the needed subset.
5. Remove stale or unused skills.

Least privilege applies to skills too.

For this personal repository, open an issue or patch with:

• affected skill or script;
• unsafe behavior;
• reproduction case;
• expected safer behavior;
• proposed fix or rollback.