diff --git a/AGENTS.md b/AGENTS.md
index e26dd3b..b072980 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -32,9 +32,10 @@ across Claude Code, Gemini CLI, Cursor, and Pi.
   market analysis, strategy, content generation, PDF render, content/design review) and
   `swiss-tech-resume-setup`. The two review skills (`resume-content-review`,
   `resume-design-review`) declare `context: fork` so they review with fresh eyes.
-- Per-agent skill directories (`.claude/skills/`, and later `.gemini/skills/`,
-  `.cursor/skills/`) are **symlinks** into `.agents/skills/`. Pi reads `.agents/skills/`
-  natively. Edit skills only in `.agents/skills/`.
+- **Pi and Cursor read `.agents/skills/` natively** (no config — just open/run in a clone).
+  **Claude Code** uses the `.claude/skills/` symlink into `.agents/skills/` (plus the plugin
+  manifest). **Gemini CLI** needs a `.gemini/skills/` symlink into `.agents/skills/` (adapter
+  pending). Either way, edit skills only in `.agents/skills/` — never through a symlink.
 - `.claude-plugin/` - Claude Code plugin + marketplace manifests (points `skills` at
   `.agents/skills/`).
 - `docs/PERSONAL_PROFILE.md` - **Primary data source** (private, gitignored). Built and
diff --git a/README.md b/README.md
index e4b16ad..c1a8c57 100644
--- a/README.md
+++ b/README.md
@@ -197,7 +197,7 @@ cd CV
 | **Claude Code** | Install as a plugin (above), or run in a clone — skills resolve via `.claude/skills/` (symlink) and the plugin manifest. |
 | **Pi** | Run `pi` inside the clone — skills under `.agents/skills/` are auto-discovered (no config). Or install remotely: `pi install git:github.com/datarian/CV` (the root `package.json` declares `pi.skills`). |
 | **Gemini CLI** | Symlink `.gemini/skills → ../.agents/skills` and point Gemini at `AGENTS.md` via `.gemini/settings.json`. *(adapter pending)* |
-| **Cursor** | Symlink `.cursor/skills → ../.agents/skills`; `AGENTS.md` is read natively. *(adapter pending)* |
+| **Cursor** | Open the clone in Cursor — `.agents/skills/` and `AGENTS.md` are both read natively (no config). |
 
 Then ask the same things you would in Claude Code (e.g. *"Build me a Swiss-market resume for
 Senior ML Engineer roles"*); invoke the orchestrator explicitly with `/skill:swiss-tech-resume-builder`
@@ -208,7 +208,9 @@ where your agent supports it.
 > configured per-agent (`.claude/.mcp.json` for Claude Code; Pi/Gemini/Cursor MCP wiring is a
 > documented follow-up). The mandatory render constraints (moderncv `fancy`, XeLaTeX-only, 6-arg
 > `\cventry`, GitHub footer) live inside the render skill itself, so they hold on every agent even
-> if it doesn't read `AGENTS.md`.
+> if it doesn't read `AGENTS.md`. The two review skills declare `context: fork` for isolation; where
+> an agent doesn't honor that field (e.g. Cursor today), they simply run in the main context — the
+> review still happens, just without a separate forked turn.
 
 ## Your data stays private