diff --git a/README.md b/README.md
index b7cdbaa..119e37e 100644
--- a/README.md
+++ b/README.md
@@ -24,6 +24,10 @@ Heavily inspired by Rust's [Iterator trait](https://doc.rust-lang.org/std/iter/t
 pip install itrx
 ```
 
+> **Using an AI coding agent?** `itrx` ships an installable [agent skill](#agent-skill) — run
+> `uv run itrx-skill --install` and your agent gets a built-in reference for writing `Itr` chains
+> correctly, without needing this whole README in context.
+
 ## Quick Start & Core Functionality
 
 ### Chaining
@@ -105,6 +109,27 @@ When working with `Itr`, keep these points in mind:
 *   **No Rewinding:** It's not possible to rewind an `Itr` to an earlier state. You can "preview" the next value using the `peek()` method, and conditionally consume it with `next_if()`.
 *   **Infinite Iterators:** Be cautious with open-ended iterators (e.g., those from `itertools.count()` or custom generators). Eager evaluation methods (like `collect()`, `count()`, `reduce()`) will attempt to consume the entire sequence, potentially leading to infinite loops or out-of-memory errors if applied to an infinite source.
 
+## Agent skill
+
+`itrx` ships a `SKILL.md` for AI coding agents (e.g. Claude Code) covering the `Itr` API, the
+lazy/eager split described above, and the common pitfalls (single-pass iterators, infinite
+sources, `groupby` vs `chunk_by`), so an agent doesn't need this whole README in context to write
+correct `Itr` chains.
+
+Install it into a project as a symlink to the version installed in the current environment:
+
+```sh
+uv run itrx-skill --install [PATH]  # default PATH: .agents
+uv run itrx-skill --remove [PATH]   # default PATH: .agents
+```
+
+(Or, without `uv`, activate the virtualenv `itrx` is installed in and drop the `uv run` prefix —
+`itrx-skill` is a normal console-script entry point, so it's only on `PATH` while that environment
+is active.)
+
+This creates (or removes) `PATH/skills/itrx`, symlinked to the skill bundled inside the installed
+`itrx` package, so it always matches the version in use.
+
 ## API Reference
 
 For a complete list of `Itr` methods and their detailed descriptions, please refer to the [API documentation](./doc/apidoc.md).
diff --git a/pyproject.toml b/pyproject.toml
index 081b916..2379c62 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -20,6 +20,9 @@ classifiers = [
 "Homepage" = "https://github.com/virgesmith/itrx"
 "Bug Tracker" = "https://github.com/virgesmith/itrx/issues"
 
+[project.scripts]
+itrx-skill = "itrx.skill_cli:main"
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
diff --git a/relnotes.md b/relnotes.md
index a8b5af8..3496152 100644
--- a/relnotes.md
+++ b/relnotes.md
@@ -1,3 +1,9 @@
+## Unreleased
+
+### New features
+
+- Installable **agent skill**: the package now bundles a `SKILL.md` reference for AI coding agents, plus an `itrx-skill` console script to symlink it into a project (`itrx-skill --install [PATH]` / `--remove [PATH]`, default `PATH=.agents`, creating `PATH/skills/itrx`). The symlink points at the skill inside the installed `itrx`, so it always matches the version in use. See the "Agent skill" section of the README.
+
 ## 0.4.0
 
 ### Breaking changes
diff --git a/src/itrx/skill/SKILL.md b/src/itrx/skill/SKILL.md
new file mode 100644
index 0000000..d80de39
--- /dev/null
+++ b/src/itrx/skill/SKILL.md
@@ -0,0 +1,168 @@
+---
+name: itrx
+description: >
+  Use when writing, editing, or debugging code that uses itrx's `Itr` — a chainable, lazy
+  iterator adaptor for Python inspired by Rust's Iterator trait. Covers the lazy/eager split,
+  the full method set by category, single-pass and infinite-iterator pitfalls, and how
+  `collect` materialises results. Triggers: "itrx", "Itr", chainable/fluent iterator, a
+  Rust-style `.map().filter().collect()` chain in Python, replacing nested `itertools` calls.
+---
+
+# Developing with itrx
+
+`itrx` wraps any Python iterable, iterator, or generator in an `Itr[T]` that exposes a fluent,
+chainable, mostly-lazy API modelled on Rust's [`Iterator`](https://doc.rust-lang.org/std/iter/trait.Iterator.html)
+trait. It is zero-dependency: in most cases each method is thin syntactic sugar over `itertools`
+or a builtin, so the runtime cost over hand-written `itertools` is negligible.
+
+```py
+from itrx import Itr
+
+Itr(range(100)).rev().step_by(4).skip(10).map(lambda x: x * x).filter(lambda x: x % 10 == 9).for_each(print)
+```
+
+`Itr` implements the iterator protocol itself (`__iter__`/`__next__`), so an `Itr` can be passed
+anywhere an iterator is accepted — `for x in itr`, `list(itr)`, `sum(itr)`, or as the `other`
+argument to `zip`/`chain` on another `Itr`.
+
+This skill is a quick reference, not the full API. For a method's exact signature, parameters,
+raises and doctest, consult the generated per-method reference:
+
+<https://github.com/virgesmith/itrx/blob/main/doc/apidoc.md>
+
+Read it whenever you need a detail this file doesn't state — it is generated from the source
+docstrings and covers every public method. Note it tracks `main`, so it may describe methods
+newer than the installed version; `Itr.<method>.__doc__` (or `help(Itr.<method>)`) is the
+authoritative check for what the *installed* version actually has. In a checkout of the itrx repo
+itself the same file is at `doc/apidoc.md`, and the narrative docs are in `README.md`.
+
+## When reaching for `Itr` is worth it
+
+The reason to use it is **readability**, not speed — it will not make a pipeline faster than the
+equivalent `itertools` code, because it *is* that code underneath. Reach for it when:
+
+- **A pipeline reads inside-out as builtins.** `filter(lambda .., map(lambda .., islice(islice(reversed(x), None, None, 4), 10, None)))`
+  has to be read from the innermost call outwards; the `Itr` chain reads left to right in the
+  order the data actually flows.
+- **You want laziness by default without writing generators.** Every intermediate step returns a
+  new `Itr` and pulls items only on demand, so an infinite source stays workable right up to the
+  terminal call.
+- **The operation exists in Rust's `Iterator` but not as a Python builtin** — `fold`, `inspect`,
+  `partition`, `position`, `intersperse`, `interleave`, `dedup`, `chunk_by`, `unzip`, `map_while`,
+  `next_chunk`, `step_by`, `rolling`.
+
+Conversely, it is **not** worth it for a single `map`/`filter` (a comprehension is clearer), for
+code already dominated by numpy/pandas vectorised calls, or where the data is a materialised
+`list` you are indexing into rather than streaming.
+
+## Constructing
+
+`Itr(it)` accepts any `Iterable[T]` and immediately calls `iter()` on it. Wrapping a **sequence**
+(list, tuple, str, range) is safe to do repeatedly — each `Itr` gets its own fresh iterator — but
+wrapping a **generator or iterator** hands over ownership: consuming the `Itr` consumes it.
+
+## Lazy vs eager — the single most important distinction
+
+**Lazy** (return a new `Itr` and consume nothing until pulled, so they are safe on infinite
+sources — with the exception of `product`, which materialises `other` up front like
+`itertools.product` does):
+
+`accumulate`, `batched`, `chain`, `chunk_by`, `copy`, `cycle`, `dedup`, `enumerate`, `filter`,
+`flat_map`, `flatten`, `inspect`, `interleave`, `intersperse`, `map`, `map_dict`, `map_while`,
+`pairwise`, `partition`, `product`, `repeat`, `rolling`, `skip`, `skip_while`, `step_by`, `take`,
+`take_while`, `tee`, `unzip`, `zip`, `zip_longest`
+
+**Eager** (consume the iterator, return a concrete value; **never** on an infinite source):
+
+- Collection: `collect`, `last`, `next`, `next_chunk`, `next_if`, `nth`, `position`, `peek`
+- Aggregation: `all`, `any`, `consume`, `count`, `find`, `fold`, `for_each`, `max`, `min`, `prod`,
+  `reduce`, `sum`
+- Whole-input reordering: `groupby`, `sorted_by`, `value_counts`, `rev`
+
+Note that some of these consume only as far as they need to: `next`, `next_chunk`, `nth`,
+`next_if`, `peek`, `find`, `position`, `any` and `all` short-circuit, so they *are* safe on an
+infinite source. `collect`, `count`, `last`, `consume`, `fold`, `reduce`, `sum`, `prod`, `max`,
+`min`, `for_each`, `rev`, `groupby`, `sorted_by` and `value_counts` are not.
+
+## Outputs
+
+`collect(container=tuple)` materialises into `tuple` (default), `list`, `set`, or `dict`:
+
+```py
+>>> from itrx import Itr
+>>> Itr(("apple", "banana", "carrot")).groupby(len).collect(dict)
+{5: ('apple',), 6: ('banana', 'carrot')}
+```
+
+- `collect(dict)` requires the iterator to yield 2-tuples — pair up with `zip`, `enumerate`,
+  `groupby`, `chunk_by`, `value_counts` or a `map` that returns pairs first.
+- `collect(set)` is the "unique" operation (and drops order).
+- `for_each(func)` is the side-effecting terminal; `consume()` exhausts for side effects alone
+  (pair it with `inspect`); `fold(init, func)` is the general terminal reduction.
+
+## Method notes and gotchas
+
+- **Single-pass, no rewinding.** Like any Python iterator, an `Itr` is consumed once. To use a
+  sequence twice, use `copy()` (splits at the current position into an independent `Itr`),
+  `tee(n)`, or re-wrap the original source. There is no `reset`.
+- **`peek()` does not advance**, and returns the same value on repeated calls; it raises
+  `StopIteration` when exhausted. `next_if(predicate)` consumes the next item **only** if it
+  satisfies the predicate, otherwise leaves it in place and returns `None` (also `None` when
+  exhausted) — the correct tool for conditional lookahead parsing.
+- **`tee(n)` and `copy()` share a buffer.** After `tee`, do not consume the original `Itr` — use
+  the returned ones. If one tee lags far behind another the buffer grows to hold the gap, so
+  memory can blow up on a large or infinite source. `tee(n)` raises `ValueError` for `n < 1`.
+- **`nth(n)` is 0-based** (like Rust's `Iterator::nth`): `nth(0)` is the first item. It raises
+  `ValueError` for `n < 0` and `StopIteration` if the iterator is shorter than `n + 1`.
+- **`next()` and `last()` raise `StopIteration`** on an empty iterator, they do not return `None`.
+  `find(predicate)` *does* return `None` when nothing matches.
+- **`groupby(key)` and `sorted_by(key)` sort the whole input up front**, so they reorder output,
+  require mutually-orderable keys, and must not touch an infinite source. Use the lazy
+  **`chunk_by(key)`** to group *consecutive* runs without sorting (the `itertools.groupby`
+  semantics) — it preserves order and works on infinite iterators.
+- **`value_counts()` is eager** and yields `(value, count)` most-common-first with ties by first
+  appearance (pandas semantics, not sorted by key). Items must be hashable. Collect it into a
+  `dict` for a lookup table.
+- **`dedup()` removes only *adjacent* duplicates**, keeping the first of each run. It compares by
+  equality (items need not be hashable) and stays lazy — it is not "unique". For global
+  uniqueness use `collect(set)`, accepting the loss of order.
+- **`rev()` materialises the entire remaining sequence** into memory before yielding — unavoidable,
+  but never call it on an unbounded source.
+- **`repeat(n)` tees the iterator `n` times**, so it buffers the whole sequence for large `n`; it
+  always returns a new `Itr` and leaves the original exhausted. `cycle()` repeats indefinitely.
+- **`flatten()` removes exactly one level** of nesting and every item must itself be iterable
+  (note that `str` is, which flattens strings into characters). `flat_map(f)` is
+  `map(f).flatten()`.
+- **`zip(other)` stops at the shorter input** (no `strict=True` mode); `zip_longest(other,
+  fillvalue=...)` runs to the longer, padding. `interleave(other)` alternates and then yields the
+  tail of whichever is longer.
+- **`starmap(f)` unpacks each item as `*args`** — use it after `zip`, `enumerate`, `pairwise` or
+  `product` instead of `map(lambda t: f(*t))`.
+- **`map_dict(mapping)`** looks each item up in a `dict` (a `defaultdict` works); a missing key
+  raises `KeyError`.
+- **`batched(n)` yields non-overlapping tuples** (the last possibly short); **`rolling(n)`** yields
+  overlapping windows of exactly `n` (a generalisation of `pairwise`); **`step_by(n)`** keeps every
+  n-th item starting with the first.
+- **`partition(predicate)`** returns `(matching, non_matching)` as two lazy `Itr`s built on a
+  `copy()`, so the predicate runs twice per item and the same buffering caveat as `tee` applies.
+- **`product(other)` materialises `other`** (as `itertools.product` does), so the *other* iterable
+  must be finite even though the chain stays lazy in `self`.
+- **`inspect(func)` is the lazy debugging hook** — it calls `func` on each item and passes it
+  through unchanged, so you can drop it mid-chain without altering results.
+
+## Typing
+
+`Itr[T]` is generic and every method is fully annotated, so type checkers track the element type
+through a chain: `Itr(range(3)).map(str)` is `Itr[str]`, `.enumerate()` is `Itr[tuple[int, str]]`,
+`.collect(list)` is `list[str]`. The package ships a `py.typed` marker. When a chain's inferred
+type is wrong, the fix is usually to annotate the lambda or the source, not to cast the `Itr`.
+
+## Working on itrx itself (not just using it)
+
+If you are editing this library's own source (`src/itrx/itr.py`) rather than using it in a
+downstream project, the repo's `AGENTS.md` governs the workflow. In short: no new runtime
+dependencies (it is zero-dependency by design), new non-terminal methods must return an `Itr` and
+stay lazy, every public method needs a doctest, the full gate suite is
+`uv run ruff check && uv run ruff format --check && uv run ty check src && uv run pytest`, and
+coverage is enforced at 100%. Regenerate `doc/apidoc.md` with
+`uv run python src/scripts/introspect.py` whenever the public API changes.
diff --git a/src/itrx/skill_cli.py b/src/itrx/skill_cli.py
new file mode 100644
index 0000000..28ca84c
--- /dev/null
+++ b/src/itrx/skill_cli.py
@@ -0,0 +1,100 @@
+import argparse
+import os
+from pathlib import Path
+
+_SKILL_NAME = "itrx"
+_DEFAULT_PATH = ".agents"
+
+
+def _source_dir() -> Path:
+    """The bundled skill directory, shipped inside this package."""
+    return Path(__file__).parent / "skill"
+
+
+def _target_path(root: str) -> Path:
+    return Path(root) / "skills" / _SKILL_NAME
+
+
+def _is_ours(target: Path, source: Path) -> bool:
+    if not target.is_symlink():
+        return False
+    try:
+        return target.resolve() == source.resolve()
+    except OSError:
+        return False
+
+
+def _link_target(source: Path, target: Path) -> str:
+    """A relative link target, or an absolute one if there is no relative path (different windows drives)."""
+    try:
+        return os.path.relpath(source.resolve(), target.parent.resolve())
+    except ValueError:
+        return str(source.resolve())
+
+
+def _install(root: str) -> int:
+    source = _source_dir()
+    target = _target_path(root)
+    target.parent.mkdir(parents=True, exist_ok=True)
+
+    if target.is_symlink():
+        if _is_ours(target, source):
+            print(f"already installed and up to date: {target}")
+            return 0
+        print(f"refusing to overwrite existing symlink not managed by {_SKILL_NAME}-skill: {target}")
+        return 1
+    if target.exists():
+        print(f"refusing to overwrite existing file or directory: {target}")
+        return 1
+
+    target.symlink_to(_link_target(source, target), target_is_directory=True)
+    print(f"installed: {target} -> {source}")
+    return 0
+
+
+def _remove(root: str) -> int:
+    source = _source_dir()
+    target = _target_path(root)
+
+    if not target.exists() and not target.is_symlink():
+        print(f"not installed: {target}")
+        return 0
+    if not _is_ours(target, source):
+        print(f"refusing to remove {target}: not a symlink managed by {_SKILL_NAME}-skill")
+        return 1
+
+    target.unlink()
+    print(f"removed: {target}")
+    return 0
+
+
+def main(argv: list[str] | None = None) -> int:
+    """Entry point for the `itrx-skill` console script."""
+    parser = argparse.ArgumentParser(
+        prog=f"{_SKILL_NAME}-skill",
+        description=f"Install or remove the '{_SKILL_NAME}' agent skill as a symlink into a project.",
+    )
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument(
+        "--install",
+        nargs="?",
+        const=_DEFAULT_PATH,
+        metavar="PATH",
+        help=f"Symlink the skill into PATH/skills/{_SKILL_NAME} (default PATH: {_DEFAULT_PATH}).",
+    )
+    group.add_argument(
+        "--remove",
+        nargs="?",
+        const=_DEFAULT_PATH,
+        metavar="PATH",
+        help=f"Remove the symlink from PATH/skills/{_SKILL_NAME} (default PATH: {_DEFAULT_PATH}).",
+    )
+    args = parser.parse_args(argv)
+
+    if args.install is not None:
+        return _install(args.install)
+    return _remove(args.remove)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())  # pragma: no cover - console script entry point
diff --git a/src/test/test_skill_cli.py b/src/test/test_skill_cli.py
new file mode 100644
index 0000000..dd37ff7
--- /dev/null
+++ b/src/test/test_skill_cli.py
@@ -0,0 +1,115 @@
+from pathlib import Path
+
+import pytest
+
+from itrx import skill_cli
+
+
+def test_install_creates_symlink(tmp_path: Path) -> None:
+    assert skill_cli.main(["--install", str(tmp_path)]) == 0
+
+    target = tmp_path / "skills" / "itrx"
+    assert target.is_symlink()
+    assert target.resolve() == skill_cli._source_dir().resolve()
+    assert (target / "SKILL.md").is_file()
+
+
+def test_install_default_path(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    monkeypatch.chdir(tmp_path)
+    assert skill_cli.main(["--install"]) == 0
+
+    target = tmp_path / ".agents" / "skills" / "itrx"
+    assert target.is_symlink()
+
+
+def test_install_across_windows_drives(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    # os.path.relpath raises when the installed package and the target project are on different drives
+    def raise_valueerror(*_args: object, **_kwargs: object) -> str:
+        raise ValueError("path is on mount 'D:', start on mount 'C:'")
+
+    monkeypatch.setattr(skill_cli.os.path, "relpath", raise_valueerror)
+    assert skill_cli.main(["--install", str(tmp_path)]) == 0
+
+    target = tmp_path / "skills" / "itrx"
+    assert target.is_symlink()
+    assert target.resolve() == skill_cli._source_dir().resolve()
+
+
+def test_is_ours_ignores_unresolvable_symlink(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    target = tmp_path / "link"
+    target.symlink_to(tmp_path, target_is_directory=True)
+
+    def raise_oserror(*_args: object, **_kwargs: object) -> Path:
+        raise OSError("cannot resolve")
+
+    monkeypatch.setattr(Path, "resolve", raise_oserror)
+    assert not skill_cli._is_ours(target, skill_cli._source_dir())
+
+
+def test_install_idempotent(tmp_path: Path) -> None:
+    assert skill_cli.main(["--install", str(tmp_path)]) == 0
+    assert skill_cli.main(["--install", str(tmp_path)]) == 0
+
+    target = tmp_path / "skills" / "itrx"
+    assert target.is_symlink()
+
+
+def test_install_refuses_existing_directory(tmp_path: Path) -> None:
+    target = tmp_path / "skills" / "itrx"
+    target.mkdir(parents=True)
+    (target / "keepme.txt").write_text("do not delete")
+
+    assert skill_cli.main(["--install", str(tmp_path)]) == 1
+    assert not target.is_symlink()
+    assert (target / "keepme.txt").read_text() == "do not delete"
+
+
+def test_install_refuses_foreign_symlink(tmp_path: Path) -> None:
+    foreign = tmp_path / "elsewhere"
+    foreign.mkdir()
+    target = tmp_path / "skills" / "itrx"
+    target.parent.mkdir(parents=True)
+    target.symlink_to(foreign, target_is_directory=True)
+
+    assert skill_cli.main(["--install", str(tmp_path)]) == 1
+    assert target.resolve() == foreign.resolve()
+
+
+def test_remove_deletes_owned_symlink(tmp_path: Path) -> None:
+    skill_cli.main(["--install", str(tmp_path)])
+    target = tmp_path / "skills" / "itrx"
+    assert target.is_symlink()
+
+    assert skill_cli.main(["--remove", str(tmp_path)]) == 0
+    assert not target.exists()
+    assert not target.is_symlink()
+
+
+def test_remove_missing_is_noop(tmp_path: Path) -> None:
+    assert skill_cli.main(["--remove", str(tmp_path)]) == 0
+
+
+def test_remove_refuses_foreign_directory(tmp_path: Path) -> None:
+    target = tmp_path / "skills" / "itrx"
+    target.mkdir(parents=True)
+    (target / "keepme.txt").write_text("do not delete")
+
+    assert skill_cli.main(["--remove", str(tmp_path)]) == 1
+    assert (target / "keepme.txt").read_text() == "do not delete"
+
+
+def test_remove_default_path(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    monkeypatch.chdir(tmp_path)
+    skill_cli.main(["--install"])
+    target = tmp_path / ".agents" / "skills" / "itrx"
+    assert target.is_symlink()
+
+    assert skill_cli.main(["--remove"]) == 0
+    assert not target.exists()
+
+
+def test_mutually_exclusive_args_required() -> None:
+    with pytest.raises(SystemExit):
+        skill_cli.main([])
+    with pytest.raises(SystemExit):
+        skill_cli.main(["--install", ".", "--remove", "."])