diff --git a/.gitignore b/.gitignore
index 7fd27471..645b6210 100644
--- a/.gitignore
+++ b/.gitignore
@@ -34,3 +34,7 @@ tests/data_*.h5
 tests/data_*/
 tests/tmp.*
 tests/.coverage
+
+# local dev artifact
+uv.lock
+.venv/
diff --git a/skills/dpdata-driver/SKILL.md b/skills/dpdata-driver/SKILL.md
new file mode 100644
index 00000000..132c36c7
--- /dev/null
+++ b/skills/dpdata-driver/SKILL.md
@@ -0,0 +1,142 @@
+---
+name: dpdata-driver
+description: Use dpdata Python Driver plugins to label systems (energies/forces/virials) via System.predict(), list available drivers, and build Driver objects (ase/deepmd/gaussian/sqm/hybrid). Use when working with dpdata Python API (not CLI) and you need driver-based energy/force prediction, plugin registration keys, or examples of using dpdata with ASE calculators or DeePMD models.
+---
+
+# dpdata-driver
+
+Use dpdata “driver plugins” to **label** a `dpdata.System` (predict energies/forces/virials) and obtain a `dpdata.LabeledSystem`.
+
+## Key idea
+
+- A **Driver** converts an unlabeled `System` into a `LabeledSystem` by computing:
+  - `energies` (required)
+  - `forces` (optional but common)
+  - `virials` (optional)
+
+In dpdata, this is exposed as:
+
+- `System.predict(*args, driver="dp", **kwargs) -> LabeledSystem`
+
+`driver` can be:
+
+- a **string key** (plugin name), e.g. `"ase"`, `"dp"`, `"gaussian"`
+- a **Driver object**, e.g. `Driver.get_driver("ase")(...)`
+
+## List supported driver keys (runtime)
+
+When unsure what drivers exist in *this* dpdata version/env, query them at runtime:
+
+```python
+from dpdata.driver import Driver
+
+print(sorted(Driver.get_drivers().keys()))
+```
+
+In the current repo state, keys include:
+
+- `ase`
+- `dp` / `deepmd` / `deepmd-kit`
+- `gaussian`
+- `sqm`
+- `hybrid`
+
+(Exact set depends on dpdata version and installed extras.)
+
+## Minimal workflow
+
+```python
+import dpdata
+from dpdata.system import System
+
+sys = System("input.xyz", fmt="xyz")
+ls = sys.predict(driver="ase", calculator=...)  # returns dpdata.LabeledSystem
+```
+
+### Verify you got a labeled system
+
+```python
+assert "energies" in ls.data
+# optional:
+# assert "forces" in ls.data
+# assert "virials" in ls.data
+```
+
+## Example: use the ASE driver with an ASE calculator (runnable)
+
+This is the easiest *fully runnable* example because it doesn’t require external QM software.
+
+Dependencies (recommended): use `uv`.
+
+Option A (one-off invocation):
+
+```bash
+uv run --with dpdata --with numpy --with ase python3 your_script.py
+```
+
+Option B (recommended for shareable scripts): declare dependencies in the script via inline metadata, then run `uv run script.py`.
+See: https://docs.astral.sh/uv/guides/scripts/#inline-metadata
+
+Script:
+
+```python
+import numpy as np
+from ase.calculators.emt import EMT
+from dpdata.system import System
+
+# write a tiny molecule
+open("tmp.xyz", "w").write("""2\n\nH 0 0 0\nH 0 0 0.74\n""")
+
+sys = System("tmp.xyz", fmt="xyz")
+ls = sys.predict(driver="ase", calculator=EMT())
+
+print("energies", np.array(ls.data["energies"]))
+print("forces shape", np.array(ls.data["forces"]).shape)
+if "virials" in ls.data:
+    print("virials shape", np.array(ls.data["virials"]).shape)
+else:
+    print("virials: <not provided by this driver/calculator>")
+```
+
+## Example: pass a Driver object instead of a string
+
+```python
+from ase.calculators.emt import EMT
+from dpdata.driver import Driver
+from dpdata.system import System
+
+sys = System("tmp.xyz", fmt="xyz")
+ase_driver = Driver.get_driver("ase")(calculator=EMT())
+ls = sys.predict(driver=ase_driver)
+```
+
+## Hybrid driver
+
+Use `driver="hybrid"` to sum energies/forces/virials from multiple drivers.
+
+The `HybridDriver` accepts `drivers=[ ... ]` where each item is either:
+
+- a `Driver` instance
+- a dict like `{"type": "sqm", ...}` (type is the driver key)
+
+Example (structure only; may require external executables):
+
+```python
+from dpdata.driver import Driver
+
+hyb = Driver.get_driver("hybrid")(
+    drivers=[
+        {"type": "sqm", "qm_theory": "DFTB3"},
+        {"type": "dp", "dp": "frozen_model.pb"},
+    ]
+)
+# ls = sys.predict(driver=hyb)
+```
+
+## Notes / gotchas
+
+- Many drivers require extra dependencies or external programs:
+  - `dp` requires `deepmd-kit` + a model file
+  - `gaussian` requires Gaussian and a valid executable (default `g16`)
+  - `sqm` requires AmberTools `sqm`
+- If you just need file format conversion, use the existing **dpdata CLI** skill instead.
diff --git a/skills/dpdata-plugin/SKILL.md b/skills/dpdata-plugin/SKILL.md
new file mode 100644
index 00000000..ef9dba2f
--- /dev/null
+++ b/skills/dpdata-plugin/SKILL.md
@@ -0,0 +1,113 @@
+---
+name: dpdata-plugin
+description: Create and install dpdata plugins (especially custom Format readers/writers) using Format.register(...) and pyproject.toml entry_points under 'dpdata.plugins'. Use when extending dpdata with new formats or distributing plugins as separate Python packages.
+---
+
+# dpdata-plugin
+
+dpdata loads plugins in two ways:
+
+1. **Built-in plugins** in `dpdata.plugins.*` (imported automatically)
+1. **External plugins** exposed via Python package entry points: `dpdata.plugins`
+
+This skill focuses on **external plugin packages**, the recommended way to add new formats without modifying dpdata itself.
+
+## What can be extended?
+
+Most commonly: add a new **Format** (file reader/writer) via:
+
+```python
+from dpdata.format import Format
+
+
+@Format.register("myfmt")
+class MyFormat(Format): ...
+```
+
+## How dpdata discovers plugins
+
+dpdata imports `dpdata.plugins` during normal use (e.g. `dpdata.system` imports it). That module:
+
+- imports every built-in module in `dpdata/plugins/*.py`
+- then loads all **entry points** in group `dpdata.plugins`
+
+So an external plugin package only needs to ensure that importing the entry-point target triggers the `@Format.register(...)` side effects.
+
+## Minimal external plugin package (based on plugin_example/)
+
+### 1) Create a new Python package
+
+Example layout:
+
+```text
+dpdata_random/
+  pyproject.toml
+  dpdata_random/
+    __init__.py
+```
+
+### 2) Implement and register your Format
+
+In `dpdata_random/__init__.py` (shortened example):
+
+```python
+from __future__ import annotations
+
+import numpy as np
+from dpdata.format import Format
+
+
+@Format.register("random")
+class RandomFormat(Format):
+    def from_system(self, N, **kwargs):
+        return {
+            "atom_numbs": [20],
+            "atom_names": ["X"],
+            "atom_types": np.zeros(20, dtype=int),
+            "cells": np.repeat(np.eye(3)[None, ...], N, axis=0) * 100.0,
+            "coords": np.random.rand(N, 20, 3) * 100.0,
+            "orig": np.zeros(3),
+            "nopbc": False,
+        }
+```
+
+Return dicts must match dpdata’s expected schema (cells/coords/atom_names/atom_types/...).
+
+### 3) Expose an entry point
+
+In `pyproject.toml`:
+
+```toml
+[project]
+name = "dpdata_random"
+version = "0.0.0"
+dependencies = ["numpy", "dpdata"]
+
+[project.entry-points.'dpdata.plugins']
+random = "dpdata_random:RandomFormat"
+```
+
+Any importable target works; this pattern points directly at the class.
+
+### 4) Install and test
+
+In a clean env (recommended via `uv`):
+
+```bash
+uv run --with dpdata --with numpy python3 - <<'PY'
+import dpdata
+from dpdata.format import Format
+
+# importing dpdata will load entry points (dpdata.plugins)
+print('random' in Format.get_formats())
+PY
+```
+
+If it prints `True`, your plugin was discovered.
+
+## Debug checklist
+
+- Did you install the plugin package into the same environment where you run dpdata?
+- Does `pyproject.toml` contain `[project.entry-points.'dpdata.plugins']`?
+- Does importing the entry point module/class execute the `@Format.register(...)` decorator?
+- If using `uv run`, remember each command runs in its own environment unless you’re in a `uv` project (or you rely on `uv run --with ...`).