diff --git a/.github/workflows/ci.yaml b/.github/workflows/ci.yaml
index 2e04727..0d3c8c8 100644
--- a/.github/workflows/ci.yaml
+++ b/.github/workflows/ci.yaml
@@ -45,7 +45,7 @@ jobs:
     - name: Install dependencies
       run: |
         # create venv in repo so it can be cached
-        python -m venv .venv
+        python -m venv --clear .venv
         source .venv/bin/activate
         pip install --upgrade pip
         uv sync --dev
diff --git a/agentflow_cli/cli/commands/skill.py b/agentflow_cli/cli/commands/skill.py
new file mode 100644
index 0000000..bdaa907
--- /dev/null
+++ b/agentflow_cli/cli/commands/skill.py
@@ -0,0 +1,86 @@
+"""Skill command — generate coding-agent skill files (CLAUDE.md, .cursorrules, ...)."""
+
+from pathlib import Path
+from typing import Any
+
+from agentflow_cli.cli.commands import BaseCommand
+from agentflow_cli.cli.exceptions import FileOperationError
+from agentflow_cli.cli.templates.skills import SKILL_TARGETS
+
+
+class SkillCommand(BaseCommand):
+    """Write AgentFlow skill files for common coding agents."""
+
+    def execute(
+        self,
+        path: str = ".",
+        agent: str = "all",
+        force: bool = False,
+        **kwargs: Any,
+    ) -> int:
+        """Generate skill files.
+
+        Args:
+            path: Project directory to generate into.
+            agent: Agent id to target. One of the keys in ``SKILL_TARGETS`` or ``"all"``.
+            force: Overwrite existing files.
+        """
+        try:
+            subtitle = (
+                "Generate AgentFlow skill files for coding agents "
+                "(Claude Code, Cursor, Copilot, Windsurf, Codex)"
+            )
+            self.output.print_banner("Skill", subtitle, color="cyan")
+
+            base_path = Path(path)
+            base_path.mkdir(parents=True, exist_ok=True)
+
+            agent_key = agent.lower().strip()
+            if agent_key == "all":
+                targets = SKILL_TARGETS
+            elif agent_key in SKILL_TARGETS:
+                targets = {agent_key: SKILL_TARGETS[agent_key]}
+            else:
+                supported = ", ".join(sorted(["all", *SKILL_TARGETS]))
+                raise FileOperationError(f"Unknown agent '{agent}'. Supported: {supported}.")
+
+            written: list[Path] = []
+            for name, (rel_path, content) in targets.items():
+                dest = base_path / rel_path
+                self._write_file(dest, content, force=force)
+                written.append(dest)
+                self.output.success(f"[{name}] wrote {dest}")
+
+            self.output.info("\n🚀 Next steps:")
+            steps = [
+                "Commit the generated files so your coding agent picks them up.",
+                "Edit the content to capture any project-specific conventions.",
+                "Re-run 'agentflow skill --force' whenever AgentFlow usage changes.",
+            ]
+            for i, step in enumerate(steps, 1):
+                self.output.info(f"{i}. {step}")
+
+            return 0
+
+        except FileOperationError as e:
+            return self.handle_error(e)
+        except Exception as e:
+            return self.handle_error(FileOperationError(f"Failed to generate skill files: {e}"))
+
+    def _write_file(self, path: Path, content: str, *, force: bool) -> None:
+        try:
+            path.parent.mkdir(parents=True, exist_ok=True)
+
+            if path.exists() and not force:
+                raise FileOperationError(
+                    f"File already exists: {path}. Use --force to overwrite.",
+                    file_path=str(path),
+                )
+
+            path.write_text(content, encoding="utf-8")
+            self.logger.debug("Wrote skill file: %s", path)
+
+        except OSError as e:
+            raise FileOperationError(
+                f"Failed to write file {path}: {e}", file_path=str(path)
+            ) from e
diff --git a/agentflow_cli/cli/main.py b/agentflow_cli/cli/main.py
index acc5b79..ff016dc 100644
--- a/agentflow_cli/cli/main.py
+++ b/agentflow_cli/cli/main.py
@@ -8,6 +8,7 @@
 from agentflow_cli.cli.commands.api import APICommand
 from agentflow_cli.cli.commands.build import BuildCommand
 from agentflow_cli.cli.commands.init import InitCommand
+from agentflow_cli.cli.commands.skill import SkillCommand
 from agentflow_cli.cli.commands.version import VersionCommand
 from agentflow_cli.cli.constants import (
     DEFAULT_CONFIG_FILE,
@@ -308,6 +309,53 @@ def build(
         sys.exit(handle_exception(e))
 
 
+@app.command()
+def skill(
+    path: str = typer.Option(
+        ".",
+        "--path",
+        "-p",
+        help="Directory to generate skill files in",
+    ),
+    agent: str = typer.Option(
+        "all",
+        "--agent",
+        "-a",
+        help=(
+            "Coding agent to target: 'claude', 'cursor', 'copilot', 'windsurf', "
+            "'codex', or 'all' (default)"
+        ),
+    ),
+    force: bool = typer.Option(
+        False,
+        "--force",
+        "-f",
+        help="Overwrite existing skill files if they exist",
+    ),
+    verbose: bool = typer.Option(
+        False,
+        "--verbose",
+        "-v",
+        help="Enable verbose logging",
+    ),
+    quiet: bool = typer.Option(
+        False,
+        "--quiet",
+        "-q",
+        help="Suppress all output except errors",
+    ),
+) -> None:
+    """Generate AgentFlow skill files for coding agents (Claude Code, Cursor, Copilot, ...)."""
+    setup_cli_logging(verbose=verbose, quiet=quiet)
+
+    try:
+        command = SkillCommand(output)
+        exit_code = command.execute(path=path, agent=agent, force=force)
+        sys.exit(exit_code)
+    except Exception as e:
+        sys.exit(handle_exception(e))
+
+
 def main() -> None:
     """Main CLI entry point."""
     try:
diff --git a/agentflow_cli/cli/templates/agentflow_skill.md b/agentflow_cli/cli/templates/agentflow_skill.md
new file mode 100644
index 0000000..436c57a
--- /dev/null
+++ b/agentflow_cli/cli/templates/agentflow_skill.md
@@ -0,0 +1,158 @@
+# AgentFlow
+
+This project uses **AgentFlow** (`10xscale-agentflow`) — a graph-based Python
+framework for building LLM agents. Use this guide when writing or modifying
+agent code in this repo.
+
+## When to use AgentFlow primitives
+
+- **`StateGraph`** — the top-level container. Nodes run, edges decide what runs next.
+- **`Agent`** — a node that calls an LLM. Handles provider SDKs, tool routing, memory, retries.
+- **`ToolNode`** — a node that executes Python functions when the LLM requests tool calls.
+- **`AgentState`** / **`Message`** — conversation state and messages flowing through the graph.
+- **`InMemoryCheckpointer`** / other checkpointers — persist state per `thread_id`.
+
+Don't reach for raw provider SDKs (`openai`, `google-genai`) inside a graph node.
+Use `Agent` with the right `provider=` instead — it handles message conversion,
+streaming, tool calls, retries, and fallbacks uniformly.
+
+## Golden-path ReAct agent
+
+```python
+from dotenv import load_dotenv
+
+from agentflow.core import Agent, StateGraph, ToolNode
+from agentflow.core.state import AgentState, Message
+from agentflow.storage.checkpointer import InMemoryCheckpointer
+from agentflow.utils.constants import END
+
+load_dotenv()
+
+
+def get_weather(location: str) -> str:
+    """Get the current weather for a location."""
+    return f"The weather in {location} is sunny"
+
+
+tool_node = ToolNode([get_weather])
+
+agent = Agent(
+    model="gpt-4o",
+    provider="openai",
+    system_prompt=[{"role": "system", "content": "You are a helpful assistant."}],
+    tool_node=tool_node,
+)
+
+
+def should_use_tools(state: AgentState) -> str:
+    last = state.context[-1] if state.context else None
+    if last and getattr(last, "tools_calls", None) and last.role == "assistant":
+        return "TOOL"
+    if last and last.role == "tool":
+        return "MAIN"
+    return END
+
+
+graph = StateGraph()
+graph.add_node("MAIN", agent)
+graph.add_node("TOOL", tool_node)
+graph.add_conditional_edges("MAIN", should_use_tools, {"TOOL": "TOOL", END: END})
+graph.add_edge("TOOL", "MAIN")
+graph.set_entry_point("MAIN")
+
+app = graph.compile(checkpointer=InMemoryCheckpointer())
+
+# Invoke with a stable thread_id to get persistent history
+res = app.invoke(
+    {"messages": [Message.text_message("What is the weather in NYC?")]},
+    config={"thread_id": "demo", "recursion_limit": 10},
+)
+```
+
+## Providers
+
+Pick a provider via the `provider=` argument on `Agent`. Model names are free-form strings.
+
+| Provider | Auth | Models |
+|---|---|---|
+| `"openai"` | `OPENAI_API_KEY` | `gpt-4o`, `gpt-4o-mini`, `o1`, `o3`, `o4-mini` |
+| `"google"` | `GEMINI_API_KEY` / `GOOGLE_API_KEY` | `gemini-2.0-flash`, `gemini-2.5-flash`, `gemini-2.5-pro` |
+| `"vertex_ai"` | `GOOGLE_CLOUD_PROJECT` + Application Default Credentials | same Gemini models |
+
+Rules:
+
+- If `provider` is omitted, it is inferred: `gpt*|o1|o3|o4` → `openai`, `gemini*` → `google`.
+- `vertex_ai` is **never** inferred — set it explicitly.
+- `google` and `vertex_ai` share model names and features; only auth differs.
+- `google-genai` is not bundled. Install it when using either Gemini path:
+  `pip install google-genai`.
+
+## Agent parameters worth knowing
+
+```python
+Agent(
+    model="gpt-4o",
+    provider="openai",
+    system_prompt=[{"role": "system", "content": "..."}],
+    tool_node=tool_node,                 # ToolNode instance or str name
+    output_type="text",                  # "text" | "json"
+    trim_context=True,                   # trim to model context window
+    reasoning_config=True,               # enable extended thinking (o1/o3/gemini-2.5)
+    retry_config=True,                   # retry transient API errors
+    fallback_models=[                    # try these if primary fails
+        "gpt-4o-mini",
+        ("gemini-2.5-flash", "google"),
+    ],
+    memory=my_memory_config,             # long-term memory retrieval
+    skills=my_skill_config,              # inject skill documents
+    multimodal_config=my_mm_config,      # auto-offload large inline media
+)
+```
+
+## State & messages
+
+- Custom state extends `AgentState`:
+
+  ```python
+  class MyState(AgentState):
+      user_id: str | None = None
+  ```
+
+- `Message.text_message("...")` builds a user message quickly.
+- Tool functions can accept `tool_call_id: str` and `state: MyState` — they are
+  injected automatically if declared.
+
+## Checkpointing
+
+- `InMemoryCheckpointer()` for dev / tests.
+- `PgCheckpointer` / `RedisCheckpointer` for production
+  (requires `REDIS_URL` or a Postgres DSN).
+- Always pass `config={"thread_id": "..."}` on `invoke` / `stream` to keep
+  history across calls.
+
+## CLI
+
+AgentFlow ships a CLI (`10xscale-agentflow-cli`):
+
+- `agentflow init` — scaffold `agentflow.json` and `graph/react.py`.
+- `agentflow api` — run the FastAPI server over your graph.
+- `agentflow play` — start the API + open the hosted playground.
+- `agentflow build` — generate a `Dockerfile` (optionally `docker-compose.yml`).
+- `agentflow skill` — regenerate these coding-agent skill files.
+
+`agentflow.json` points at the compiled graph (`"agent": "graph.react:app"`).
+
+## Don't
+
+- Don't call provider SDKs directly from a graph node. Use `Agent`.
+- Don't mutate `state` in place — return a new state/messages from your node.
+- Don't infer `vertex_ai` from the model name. Pass it explicitly.
+- Don't commit API keys; keep them in `.env` (referenced by `agentflow.json`).
+- Don't hand-roll retry loops around `Agent` — use `retry_config=True`.
+
+## Where to look
+
+- Python reference: `agentflow.core`, `agentflow.core.graph`, `agentflow.core.state`,
+  `agentflow.storage.checkpointer`, `agentflow.utils.constants`.
+- Runtime adapters (`GoogleGenAIConverter`, OpenAI converters) live under
+  `agentflow.runtime.adapters.llm` — only needed when wrapping raw SDK calls.
diff --git a/agentflow_cli/cli/templates/skills.py b/agentflow_cli/cli/templates/skills.py
new file mode 100644
index 0000000..e0cd7ee
--- /dev/null
+++ b/agentflow_cli/cli/templates/skills.py
@@ -0,0 +1,32 @@
+"""Skill templates for coding agents (Claude Code, Cursor, Copilot, Windsurf, Codex).
+
+The shared body lives in ``agentflow_skill.md`` so the rendered markdown can
+remain well-formatted without fighting Python line-length rules.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Final
+
+
+_TEMPLATE_DIR: Final[Path] = Path(__file__).parent
+_SKILL_BODY: Final[str] = (_TEMPLATE_DIR / "agentflow_skill.md").read_text(encoding="utf-8")
+
+
+_CURSOR_FRONTMATTER: Final[str] = (
+    "---\n"
+    "description: AgentFlow framework conventions and golden-path snippets\n"
+    "alwaysApply: true\n"
+    "---\n\n"
+)
+
+
+# Registry mapping agent id -> (relative output path, file content).
+SKILL_TARGETS: Final[dict[str, tuple[str, str]]] = {
+    "claude": ("CLAUDE.md", _SKILL_BODY),
+    "cursor": (".cursor/rules/agentflow.mdc", _CURSOR_FRONTMATTER + _SKILL_BODY),
+    "copilot": (".github/copilot-instructions.md", _SKILL_BODY),
+    "windsurf": (".windsurfrules", _SKILL_BODY),
+    "codex": ("AGENTS.md", _SKILL_BODY),
+}
diff --git a/tests/cli/test_skill_command.py b/tests/cli/test_skill_command.py
new file mode 100644
index 0000000..d8428e0
--- /dev/null
+++ b/tests/cli/test_skill_command.py
@@ -0,0 +1,85 @@
+"""Tests for `agentflow skill` command."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from typer.testing import CliRunner
+
+import agentflow_cli.cli.main as main_mod
+from agentflow_cli.cli.templates.skills import SKILL_TARGETS
+
+
+runner = CliRunner()
+
+
+def test_skill_all_writes_every_target(tmp_path: Path) -> None:
+    result = runner.invoke(main_mod.app, ["skill", "--path", str(tmp_path)])
+
+    assert result.exit_code == 0, result.output
+    for rel_path, _ in SKILL_TARGETS.values():
+        assert (tmp_path / rel_path).exists()
+
+    claude = (tmp_path / "CLAUDE.md").read_text(encoding="utf-8")
+    assert "# AgentFlow" in claude
+    assert "provider=" in claude
+
+
+def test_skill_single_agent_writes_only_that_file(tmp_path: Path) -> None:
+    result = runner.invoke(
+        main_mod.app,
+        ["skill", "--path", str(tmp_path), "--agent", "cursor"],
+    )
+
+    assert result.exit_code == 0, result.output
+    assert (tmp_path / ".cursor" / "rules" / "agentflow.mdc").exists()
+    assert not (tmp_path / "CLAUDE.md").exists()
+    assert not (tmp_path / "AGENTS.md").exists()
+
+
+def test_skill_cursor_has_frontmatter(tmp_path: Path) -> None:
+    result = runner.invoke(
+        main_mod.app,
+        ["skill", "--path", str(tmp_path), "--agent", "cursor"],
+    )
+    assert result.exit_code == 0, result.output
+
+    content = (tmp_path / ".cursor" / "rules" / "agentflow.mdc").read_text(encoding="utf-8")
+    assert content.startswith("---\n")
+    assert "alwaysApply: true" in content
+
+
+def test_skill_refuses_existing_file_without_force(tmp_path: Path) -> None:
+    (tmp_path / "CLAUDE.md").write_text("existing", encoding="utf-8")
+
+    result = runner.invoke(
+        main_mod.app,
+        ["skill", "--path", str(tmp_path), "--agent", "claude"],
+    )
+
+    assert result.exit_code != 0
+    assert (tmp_path / "CLAUDE.md").read_text(encoding="utf-8") == "existing"
+
+
+def test_skill_force_overwrites(tmp_path: Path) -> None:
+    (tmp_path / "CLAUDE.md").write_text("existing", encoding="utf-8")
+
+    result = runner.invoke(
+        main_mod.app,
+        ["skill", "--path", str(tmp_path), "--agent", "claude", "--force"],
+    )
+
+    assert result.exit_code == 0, result.output
+    new_content = (tmp_path / "CLAUDE.md").read_text(encoding="utf-8")
+    assert new_content != "existing"
+    assert "# AgentFlow" in new_content
+
+
+def test_skill_unknown_agent_fails(tmp_path: Path) -> None:
+    result = runner.invoke(
+        main_mod.app,
+        ["skill", "--path", str(tmp_path), "--agent", "does-not-exist"],
+    )
+
+    assert result.exit_code != 0
+    assert "Unknown agent" in result.output