Claude wrote a secret to the wrong location despite explicit, unambiguous instructions

Asked Claude to create a database-password secret in two GitHub repos (`learning-ops` and `learning-infra`), specifying it needed to exist as a GitHub Actions environment secret in both `staging` and `production` environments on both repos, with matching values per environment.

Claude correctly created the GitHub secrets on `learning-ops`, but for `learning-infra` it instead wrote the value directly into Azure Key Vault — a different system, using a pattern that didn't match how every comparable secret in that codebase already works (as a GitHub environment secret consumed at deploy time, not a directly-written Key Vault value). The instruction named the exact secret name and exact two repos; there wasn't real ambiguity to misread.

When corrected, Claude fixed it accurately and cleaned up the incorrect Key Vault secrets on request, but the initial error should not have happened given how explicit the original instructions were.

No file, test, or entry point is named. Start by reproducing the request across learning-ops and learning-infra, then trace how Claude selects between GitHub environment secrets and Azure Key Vault. Done means the requested values are created only in the specified staging and production GitHub environments for both repositories, with no unintended Key Vault secrets.

索引モデルが issue の本文から書いたものです。