diff --git a/.claude/skills/composable-pipelines/SKILL.md b/.claude/skills/composable-pipelines/SKILL.md
new file mode 100644
index 0000000..5fc1605
--- /dev/null
+++ b/.claude/skills/composable-pipelines/SKILL.md
@@ -0,0 +1,126 @@
+---
+name: composable-pipelines
+description: Build AI pipelines with the ComposablePipelines Swift package — author a declarative Pipeline (result-builder DSL + @State), compile it to a graph, and run it through an Executor. Use when writing or debugging ComposablePipelines code: Pipeline/Model/Guardrail/While/ForEach/ClientTask, @State data flow, the Executor seam, reactive While/branching, tool-calling agent loops, or when a file imports ComposablePipelines / PipelineDSL / PipelineCompiler / ExecutionEngine.
+---
+
+# Composable Pipelines
+
+A runtime-agnostic stack for AI pipelines in Swift: a declarative **DSL** lowers to a `Codable`
+**AST**, a **compiler** emits an execution graph, and an observable **walker** runs it — while an
+`Executor` you supply decides what "run a model" means.
+
+## Quick start
+
+```swift
+import ComposablePipelines
+
+struct Summary: Pipeline {
+    typealias Output = String
+    let document: String
+
+    @State var keyPoints = ""
+    @State var summary = ""
+
+    var body: some Pipeline {
+        $keyPoints.set {                                    // step 1 → slot
+            Model<String>().systemPrompt("Extract the 5 key points.").message(document)
+        }
+        $summary.set {                                      // step 2 reads step 1
+            Model<String>().systemPrompt("Summarize from these points.").input { $keyPoints.get() }
+        }
+        $summary.get()                                      // pipeline output
+    }
+}
+
+// Run it (linear flow): compile → walk.
+let graph  = PipelineCompiler().compile(pipeline.loweredGraph())
+let result = try await PipelineWalker(executor: MyExecutor()).run(graph: graph) { event in print(event) }
+let text   = try JSONDecoder().decode(String.self, from: result)   // ExecutionValue == Data (JSON)
+```
+
+## Authoring cheatsheet
+
+- **Pipeline**: `struct X: Pipeline { typealias Output = T; @State var … ; var body: some Pipeline { … } }`.
+  The graph is inferred from the `@State` slots each step reads/writes — never wired by hand.
+- **State**: `@State var s = ""` → write with `$s.set { <pipeline> }` or `$s.set(value)`; read with
+  `$s.get()`. End `body` with the output slot's `.get()`.
+- **Model** (output-only generic): `Model<Output>()` then chain `.systemPrompt(_)`,
+  `.input { $slot.get() }` **or** `.message(staticValue)`, and optionally `.tools([…])`,
+  `.temperature(_)`, `.maxTokens(_)`, `.requirements(ModelSelectionRequirements(traits: […]))`.
+- **Primitives**: `Guardrail(input, rules:allowed:blocked:)`, `While(condition:) { … }`,
+  `Group { … }`, `ForEach(in: xs) { x in … }`, `Summarize(text: $slot, maxTokens:)`,
+  `ClientTask(input: $slot) { value in … }`, `From(provider, query: $slot)`, `Self.return(value)`.
+- **Control flow**: use native `if` / `switch` on produced state (e.g. `if verdict == "trivial"`).
+
+## Running a pipeline
+
+- **Linear flow** (sequential, `ForEach`, retrieval, runtime `.get()`): compile once, then
+  `PipelineWalker(executor:).run(graph:)`.
+- **Reactive flow** (a `While` loop, or `if`/`switch` that branch on a *produced* value): use
+  `PipelineRunner.run(pipeline, executor:)`. It re-lowers and recompiles after each committed write
+  so the condition/branch re-reads state. A single `walker.run` is **not** enough for these.
+
+## The Executor seam
+
+The walker never calls a model itself. Implement one method:
+
+```swift
+struct MyExecutor: Executor {
+    func runModel(config: ModelConfig, arguments: ModelArguments,
+                  onDelta: (@Sendable (String) -> Void)?) async throws -> ExecutionValue {
+        let system = arguments.systemPrompt                         // String
+        let user: String; if case .string(let s)? = arguments.message { user = s } else { user = "" }
+        // … call your model … then encode the output as JSON Data:
+        return try JSONEncoder().encode(reply)                       // or ModelTurn for tool turns
+    }
+}
+```
+
+`MockExecutor` ships for tests; `OpenAIChatExecutor` (Foundation-only) talks to any
+OpenAI-compatible endpoint (local or remote).
+
+## Tool-calling agents
+
+A model step whose output is `ModelTurn` can request tools; you dispatch them and loop:
+
+```swift
+While(condition: { reply.isEmpty && turns < maxTurns }) {
+    $lastTurn.set { Model<ModelTurn>().tools(tools.map(\.descriptor)).systemPrompt(prompt).input { $transcript.get() } }
+    $transcript.set {
+        ClientTask(input: $lastTurn) { turn in
+            guard let calls = turn.toolCalls, !calls.isEmpty else { return transcript }
+            var t = transcript
+            for call in calls {
+                let out = try await ToolRegistry(tools).executeJSON(toolName: call.name, inputJSON: Data(call.arguments.utf8))
+                t += "\n[\(call.name)] " + String(decoding: out, as: UTF8.self)
+            }
+            return t
+        }
+    }
+    $reply.set { ClientTask(input: $lastTurn) { $0.toolCalls?.isEmpty == false ? "" : ($0.reply ?? "") } }
+    $turns.set { ClientTask(input: $turns) { $0 + 1 } }
+}
+$reply.get()
+```
+
+Tools are `ModelTool`s (static `name`/`description`/`inputSchema` + `func call(_:) -> Output`),
+registered in a `ToolRegistry`. Drive tool/agent loops with `PipelineRunner.run`.
+
+## Pitfalls (read before writing code)
+
+- **`Model` takes one generic — the output.** It's `Model<Output>().systemPrompt(…).input { … }`,
+  **not** `Model<Input, Output>(instructions:input:)` (that API does not exist).
+- **Never read `@State` inside a `ClientTask` closure.** Thread state in via the typed input:
+  `ClientTask(input: $slot) { value in … }`. Reading `@State` in the closure is disallowed.
+- **`While` / value-dependent branching need `PipelineRunner.run`**, not a bare `walker.run` — they
+  require reactive re-lowering.
+- **`ExecutionValue` is JSON `Data`.** Decode results (`JSONDecoder().decode(T.self, from:)`) and
+  encode executor outputs (`JSONEncoder().encode(_)`).
+- **A turn can carry a preamble *and* tool calls.** In an agent loop, keep looping while there are
+  tool calls; only a tool-free turn is the final answer.
+
+## Reference
+
+Deep docs live in the package: [`docs/examples.md`](../../../docs/examples.md) (worked pipelines
+with code), `docs/primitives.md`, `docs/executors.md`, `docs/architecture.md`, and the top-level
+`README.md`. The runnable examples in `Examples/` are execution-tested — copy from them.
diff --git a/.claude/skills/install.sh b/.claude/skills/install.sh
new file mode 100755
index 0000000..ee6c78f
--- /dev/null
+++ b/.claude/skills/install.sh
@@ -0,0 +1,55 @@
+#!/bin/sh
+#
+# Install the ComposablePipelines agent skills into the skill directories of major AI coding agents.
+#
+# Usage:
+#   curl -fsSL https://raw.githubusercontent.com/MacPaw/ComposablePipelines/main/.claude/skills/install.sh | sh
+#
+# Installs every skill bundled in this repo (each directory under .claude/skills/ with a SKILL.md),
+# so new skills are picked up automatically. By default it installs for Claude Code and for any other
+# supported agent whose config directory already exists. Override the destinations explicitly:
+#   SKILL_DIRS="$HOME/.claude/skills $HOME/.codex/skills"   # space-separated
+# Other env: CP_SKILLS_REPO (repo URL), CP_SKILLS_BRANCH (branch).
+#
+set -eu
+
+REPO="${CP_SKILLS_REPO:-https://github.com/MacPaw/ComposablePipelines}"
+BRANCH="${CP_SKILLS_BRANCH:-main}"
+
+command -v git >/dev/null 2>&1 || { echo "error: git is required to install the skills" >&2; exit 1; }
+
+# Destinations: explicit override, or Claude Code + any other agent whose home dir is present.
+if [ -n "${SKILL_DIRS:-}" ]; then
+    targets="$SKILL_DIRS"
+else
+    targets="$HOME/.claude/skills"                                       # Claude Code (always)
+    [ -d "$HOME/.codex" ]           && targets="$targets $HOME/.codex/skills"            # Codex CLI
+    [ -d "$HOME/.config/opencode" ] && targets="$targets $HOME/.config/opencode/skills"  # opencode
+    [ -d "$HOME/.gemini" ]          && targets="$targets $HOME/.gemini/skills"           # Gemini CLI
+    [ -d "$HOME/.copilot" ]         && targets="$targets $HOME/.copilot/skills"          # Copilot CLI
+fi
+
+tmp="$(mktemp -d)"
+trap 'rm -rf "$tmp"' EXIT
+
+echo "Fetching skills from $REPO ($BRANCH)…"
+git clone --depth 1 --branch "$BRANCH" "$REPO" "$tmp/repo" >/dev/null 2>&1
+src="$tmp/repo/.claude/skills"
+
+total=0
+for dest in $targets; do
+    mkdir -p "$dest"
+    n=0
+    for dir in "$src/"*/; do
+        [ -f "${dir}SKILL.md" ] || continue          # only real skill directories
+        name="$(basename "$dir")"
+        rm -rf "$dest/$name"
+        cp -R "$dir" "$dest/$name"
+        n=$((n + 1))
+    done
+    echo "  $dest  ($n skill(s))"
+    total=$((total + n))
+done
+
+[ "$total" -gt 0 ] || { echo "error: no skills found under .claude/skills/ in $REPO" >&2; exit 1; }
+echo "Done. Restart your agent to pick up the skills."
diff --git a/README.md b/README.md
index 22f3db7..184ab26 100644
--- a/README.md
+++ b/README.md
@@ -16,6 +16,11 @@ the data dependencies and emits an execution graph; and an observable interprete
 — runs that graph with incremental, epoch-based re-execution, delegating every model and tool step
 to a backend you supply through a single `Executor` interface.
 
+The same primitives scale from a two-step summary to a full **tool-calling agent**: loops, tools,
+and branching are ordinary pipeline constructs. See the
+[coding-agent worked example](#worked-example-a-coding-agent-as-a-pipeline) — a real agent built
+entirely on this DSL.
+
 ```swift
 import ComposablePipelines
 
@@ -296,6 +301,20 @@ struct CodingAgentPipeline: Pipeline {
 > are rejected). `bash` is an escape hatch — it runs with the directory as its cwd but is not
 > otherwise sandboxed, so point the agent at a scratch directory and a model you trust.
 
+## Build pipelines with your AI agent
+
+An agent **skill** aggregates the API, patterns, and common pitfalls of building pipelines, so your
+AI coding agent authors and debugs them correctly. Install it for the major agents:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/MacPaw/ComposablePipelines/main/.claude/skills/install.sh | sh
+```
+
+The script installs every bundled skill into the skill directory of each supported agent it finds
+(Claude Code, Codex, opencode, Gemini CLI, Copilot CLI); set `SKILL_DIRS` to choose destinations.
+Each skill activates automatically when you work in a project that uses ComposablePipelines. Source:
+[`.claude/skills/`](.claude/skills/).
+
 ## Documentation
 
 - [Getting started](docs/getting-started.md) — install, your first pipeline, compile + walk.