diff --git a/dev/README_RELEASE_PROVIDERS.md b/dev/README_RELEASE_PROVIDERS.md
index 04b25799e9c86..36841800f72e9 100644
--- a/dev/README_RELEASE_PROVIDERS.md
+++ b/dev/README_RELEASE_PROVIDERS.md
@@ -196,7 +196,7 @@ First thing that release manager has to do is to convert commits for each provid
 and update version of the provider to a target version - depending on type of changes implemented in the
 providers.
 
-The recommended way to do this is via the **`prepare-providers-documentation` skill** loaded by an
+One option for doing this is the **`prepare-providers-documentation` skill** loaded by an
 agentic coding framework (e.g. Claude Code or OpenAI Codex CLI), which replaces the manual
 commit-by-commit classification step of `breeze release-management
 prepare-provider-documentation` with AI-driven classification. The skill inspects every PR (using
@@ -204,6 +204,24 @@ sub-agents per PR for thorough analysis), pays special attention to potentially
 reading the actual diff (not just the commit message or PR labels), scopes multi-provider PRs to the
 slice that touched the current provider, and asks the release manager only when genuinely uncertain.
 
+> [!WARNING]
+> The skill is an **optional** aid, not a required or default part of the release process. The
+> interactive `breeze release-management prepare-provider-documentation` command (see
+> [Falling back to interactive breeze](#falling-back-to-interactive-breeze)) remains the baseline,
+> and release managers who do not use AI tooling should use that command directly — this skill is
+> not for them. If you do use the skill, keep the following in mind:
+>
+> * **You must verify its output.** The version bumps and changelog entries it produces are
+>   generated by an LLM and are **not authoritative** — the release manager is responsible for
+>   reviewing every generated entry (especially breaking-change classification and version bumps)
+>   before merging the release PR. Treat the skill's output as a first draft to be checked, not a
+>   final answer.
+> * **It runs for a long time and consumes a lot of tokens.** Classifying a full release wave with
+>   per-PR sub-agents can take a while and burn through a large amount of model tokens. Because of
+>   that, it is hard to recommend without a high-capacity plan (e.g. Claude Max, or an equivalent
+>   GitHub Copilot / OpenAI tier). On a metered or low-quota plan the cost may be significant and
+>   the run may not complete.
+
 **Prerequisites:** running this skill requires an agentic coding environment with the
 [GitHub MCP server](https://github.com/github/github-mcp-server) configured — the skill reads PR
 diffs, lists commits, and (with maintainer confirmation) edits files via these tools. Two well-known