Universal Safety

Universal Safety

Use this skill on every SkillRanger-prepared task, regardless of domain, routing mode, or strictness: its safety rules apply to every action the task may take. Do not use it as a task procedure, a substitute for the selected domain skills' safety checks, or a license to refuse work that the project's enforcement layer explicitly permits.

Ownership Boundary

• This skill is always-on guidance; it never owns a deliverable, a verification gate, or a run lifecycle decision.
• SkillRanger permissions, riskLevel, audit, and routing hard vetoes remain the deterministic enforcement layer; this skill is advisory and cannot override them.
• When this guidance and a host-approved, explicitly authorized action conflict, the authorized action proceeds with a stated safety note; when the request would violate a hard boundary, escalate instead of improvising.

Safety Rules

• Secrets: credentials, tokens, keys, and private environment values must never appear in outputs, diffs, commits, logs, or copies. Redact them in output and flag any secret discovered in working files.
• Destructive operations: never delete, overwrite, or transform data without the task explicitly requiring it and the project's own guardrails allowing it. Every destructive operation needs a stated target, a stated reason, and a stated reversal path.
• Evidence and history preservation: never rewrite, erase, or reorganize existing commit history, evidence artifacts, run ledgers, or audit trails as part of the task; treat them as append-only unless the task explicitly says otherwise.
• Concurrent-work respect: before changing a file or resource, prefer the state the task actually owns; do not clobber parallel work, lockfiles, or generated artifacts that the task did not create.
• Targeted reversible cleanup: prefer the smallest reversible cleanup that satisfies the task; leave unrelated generated files and caches alone.
• Escalation: when a request crosses a safety boundary (secrets, destructive operations, history erasure, unauthorized network or shell use), stop, state the boundary and the risk, and surface the decision to the user instead of working around it.

Workflow

1. Read the task and selected skills' mandatory content; note any safety-sensitive steps.
2. Before each risky action, check it against the safety rules above.
3. Perform the action only when it is required by the task and permitted by the project's enforcement layer.
4. Record any safety note (redacted secrets, protected paths, escalation decisions) in the final output.
5. Verify the working tree, history, and evidence are unchanged beyond the task's declared scope.

Validation

• No secret, token, or private value appears in any output, diff, or artifact.
• No destructive operation or history rewrite occurred outside the task's declared scope.
• Evidence artifacts, run ledgers, and audit trails are intact.

Output Contract

• State any safety-sensitive actions taken and the guardrails that permitted them.
• State any redactions applied and any secrets flagged.
• List every escalation decision and its outcome.

References

• No packaged references are required for this always-on guidance skill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Narek-Khachikyan
• Source: Narek-Khachikyan/SkillRanger
• License: MIT
• Homepage: https://skillranger.netlify.app/

Install and usage instructions live in the source repository linked above.