diff --git a/CLAUDE.md b/CLAUDE.md
index 189c03e..89fa9d4 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -9,7 +9,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 - **Mark tasks in td**: `td start <id>` when beginning, `td close <id>` when done.
 - **Run td commands one at a time**: Never chain `td` commands with `&&` or `;`. Each `td` call must be its own separate shell invocation.
 - **Only the orchestrator touches td**: Sub-agents must never call `td` commands directly. Concurrent `td` writes from parallel agents corrupt the SQLite database. The orchestrator handles all `td start`/`td close` calls before and after delegating to sub-agents.
-- **td approve requires a sub-agent**: `td approve` cannot be run by the same agent that implemented the task. Always spin up a sub-agent to review and run `td approve <id>` after verifying the work.
+- **td approve requires a separate session**: `td approve` cannot be run by the implementing agent or any sub-agent spawned from the same session — `td` tracks session involvement. Approval must come from a new conversation/session. Leave implemented tasks in review state (`td review <id>`) at the end of the session; they can be approved at the start of a fresh session.
 
 ## Task Management