AGENTS.md / CLAUDE.md instructions

This is the generic version, for any coding agent that reads repository-level instructions; Claude Code and Codex have their own pages with worked examples, but the underlying instruction is the same everywhere.

The idea

Wirecheck supplies evidence. The agent decides what to do with it, never the other way around. There is no safe_to_remove field anywhere in the API; an agent that sees “0 known dependants” with “2 unknown relationships” should not treat that as a green light.

Minimal instruction block

AGENTS.md or CLAUDE.md
Before removing or modifying infrastructure dependencies, use Wirecheck to
inspect impact and unresolved relationships. After the change, compare
state and verify that intended dependencies were actually removed.

Full instruction block

Paste this version if your agent can execute multi-step instructions reliably:

AGENTS.md or CLAUDE.md
## Before changing AI/integration dependencies

This repository is tracked by Wirecheck. Before modifying or removing an
existing credential, MCP server, webhook, scheduled job, external API
integration, or agent/model integration:

1. Run `wirecheck preflight <name> --json` against the target.
2. Read `direct_dependants`, `indirect_dependants`, and
   `unknown_relationships`. A component with zero known dependants but one
   or more unknown relationships is NOT confirmed safe to remove; say so
   explicitly rather than proceeding as if it were.
3. Inspect `evidence` (or run `wirecheck evidence <name> --json`) before
   claiming to understand why the dependency exists.
4. Make the requested change.
5. Report the known and unknown dependants you found, and whether any of
   them still need updating, before considering the task complete.
6. If verification is needed after a merge and rescan, run
   `wirecheck verify retirement <name> --json` and report its `status`
   (`REMOVAL_VERIFIED`, `UNRESOLVED_REMAIN`, or `CONSUMERS_REMAIN`)
   rather than asserting the removal succeeded on your own.

See the CLI reference for every command these instructions call, or the Claude Code and Codex guides for worked examples.