# Side Effect Safety

> Use immediately before an authorized write, delete, publish, migration, replacement, retry, or external side effect. Verify target, blast radius, rollback, idempotency, and post-change state. Do not trigger for reads, planning, or reversible local edits.

- Skill: `xiongweilin/side-effect-safety` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add xiongweilin/side-effect-safety`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xiongweilin/side-effect-safety/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: xiongweilin (https://skillmd.com/u/xiongweilin)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/xiongweilin/side-effect-safety

---


# Side-Effect Safety

The checks below are mandatory safety controls for authorized side effects, not optional post-implementation validation. They remain required when general validation is user-gated.

Before the effect:

1. Resolve the exact target and scope with a read-only check.
2. Identify blast radius, partial-failure modes, reversibility, and a rollback or compensation path.
3. Make retries idempotent or explicitly bounded.
4. Split destructive work into independently verifiable steps.

For replacement or cleanup, verify the new backup or replacement before retiring the old one.

After the effect, verify the resulting state and leave failures visible. Do not treat a model's completion claim as evidence.

## Success signal

A reviewer can identify the target, blast radius, rollback or compensation path, idempotency guard, and fresh post-change verification.