Ax Playbook

---
name: ax-playbook
description: This skill helps an LLM generate correct playbook code using @ax-llm/ax. Use when the user asks about playbook(), AxPlaybook, context playbooks, evolving context, ACE / Agentic Context Engineering, agent.playbook(), or growing/applying task knowledge offline and online with evolve() and update().
version: "25.1.0"
---

# Playbook Codegen Rules (@ax-llm/ax)

Use this skill to generate context-playbook code. A playbook grows an evolving body of task knowledge and renders it into a program's context. The evolution engine (ACE — Agentic Context Engineering) is hidden behind `playbook(...)`, exactly as `optimize(...)` hides its optimizer. Prefer the `playbook(...)` concept; only reach for `AxACE` directly when the user explicitly wants the low-level engine.

## Use These Defaults

- Create with `playbook(program, { studentAI, teacherAI? })`; it returns an `AxPlaybook` handle.
- Grow offline with `await pb.evolve(examples, metric)` — returns `{ bestScore, playbook }`.
- Grow online with `await pb.update({ example, prediction, feedback })` — no metric needed.
- Apply with `pb.applyTo(program)` (defaults to the bound program).
- Persist with `pb.toJSON()` and restore with `playbook(program, opts).load(snapshot)`.
- Inspect with `pb.render()` (markdown) and `pb.getState()` (`{ playbook, artifact }`).
- For agents use `agent.playbook({ target: 'actor' | 'responder' })`; default target is `'actor'`.
- Use a cheaper `studentAI` to run the program and an optional stronger `teacherAI` to reflect/curate.
- Pass AI service options for the reflect/curate calls as `teacherOptions`. A teacher marked `isExpensive` (for example OpenAI `gpt-5.5-pro`) is rejected unless you pass `teacherOptions: { useExpensiveModel: 'yes' }`. The generated Python, Java, C++, Go, and Rust packages apply the same gate and accept `teacherOptions` in their option maps.
- Prefer `ai()`, `ax()`, and `agent()` for new code.

## Critical Rules

- `playbook(...)` binds to an `AxGen` program; `evolve`/`update` need that program's signature.
- `evolve()` returns only `{ bestScore, playbook }`. There is no Pareto front and no `optimizedProgram` — that is `optimize(...)`'s shape, not a playbook's.
- `update({ example, prediction, feedback })` requires the full `{ example, prediction }`; `example` must match the program's input fields (plus any expected output). Do not pass bare input fields at the top level.
- `update()` works without a prior `evolve()`/`load()` — the handle hydrates lazily on first use.
- `applyTo()` injects a `## Context Playbook` block into the program description; calling it repeatedly recomposes from the original base (no stacking).
- Keep the offline `metric` deterministic and cheap, like a GEPA metric.
- A playbook is plain JSON. Persist `pb.toJSON()` and `load(...)` it into a fresh program for production.
- The playbook engine, construction-time agent attachment, failure harvesting,
  and verified agent evolution are available in TypeScript and the generated
  Python, Java, C++, Go, and Rust packages. Use each package's native casing and
  callback types.

## Offline Pattern (evolve)

```typescript
import { type AxMetricFn, ai, ax, playbook } from '@ax-llm/ax';

const program = ax('review:string -> sentiment:class "positive, negative"');
const studentAI = ai({ name: 'openai', apiKey: process.env.OPENAI_APIKEY! });
const metric: AxMetricFn = ({ prediction, example }) =>
  (prediction as any).sentiment === (example as any).sentiment ? 1 : 0;

const pb = playbook(program, { studentAI, maxEpochs: 2 });
const { bestScore } = await pb.evolve(train, metric);
pb.applyTo(program);
```

## Online Pattern (update)

```typescript
// After a real run, feed the outcome back so the playbook keeps learning.
await pb.update({
  example: { review: 'Five stars, would buy again.' },
  prediction: { sentiment: 'negative' },
  feedback: 'WRONG: enthusiastic praise is positive.',
});
pb.applyTo(program);
```

## Persist And Restore

```typescript
const snapshot = pb.toJSON(); // { playbook, artifact } — plain JSON
// later, in another process / a production program instance:
playbook(prodProgram, { studentAI }).load(snapshot).applyTo(prodProgram);
```

## Agents

`a.playbook({ target })` returns an agent-aware `AxAgentPlaybook` (the stage `AxPlaybook` handle plus an agent-level `evolve`). The one playbook the agent renders into its prompt grows three ways:

- Continuous (trust): the construction-time `playbook` option (see `ax-agent`) harvests each run's failures automatically — no dataset.
- On-demand (trust): `apb.update({ example, prediction, feedback })`.
- Batch verified (proof): `apb.evolve(dataset, options)` runs the full agent over a task set, mines failure clusters, and proposes one playbook bullet per weakness; with `verify` (default on) it keeps a bullet only if held-in improves AND the `validation` held-out set does not regress, else exact rollback. `verify: false` = trust-batch. Bullets-only.

```typescript
const a = agent('ticket:string -> reply:string', { ai });
const apb = a.playbook({ target: 'actor' }); // agent-aware handle; 'actor' (default) or 'responder'
await apb.update({ example, prediction, feedback }); // online: injected into the live stage prompt
const result = await apb.evolve(
  { train, validation }, // AxAgentEvalDataset
  { metric, runsPerTask: 2 }, // verify:true by default
);
```

Each teacher call uses the `teacherOptions` set next to its model:

- `a.playbook({ teacherAI, teacherOptions })` and the construction-time `playbook: { teacherAI, teacherOptions }` config cover the reflector and curator calls, including the ones `apb.evolve(...)` makes to apply a proposal.
- `apb.evolve(dataset, { teacherAI, teacherOptions })` covers the weakness miner.
- Both teachers default to the agent's `judgeAI`, then the student. With an `isExpensive` `judgeAI`, set `teacherOptions: { useExpensiveModel: 'yes' }` in both places, plus `judgeOptions` when the built-in judge scores the runs.

The agent-level `evolve(dataset, options)` is distinct from the program-level `pb.evolve(examples, metric)` above: it takes an `AxAgentEvalDataset` plus options, runs the whole pipeline, and returns baseline/final held-in & held-out with per-bullet outcomes (no `{ bestScore }`). For full-pipeline tuning of agent instructions and demos (not the playbook) use `agent.optimize(...)` (GEPA).

Generated packages expose that same agent-bound loop with language-shaped APIs:

| Language | Agent-bound evolve call |
|---|---|
| Python | `agent.playbook().evolve(dataset, options)` |
| Java | `agent.playbook(null).evolve(dataset, options)` |
| C++ | `agent.get_playbook()->evolve(dataset, options)` |
| Go | `agent.GetPlaybook().EvolveAgent(ctx, dataset, options)` |
| Rust | `playbook.evolve_agent(&mut agent, client, dataset, options)` |

All five generated packages thread structured `failureSignals` through agent
evaluation predictions. The default verify gate accepts a proposed bullet only
when held-in score improves and held-out score stays within `epsilon`; rejection
restores the exact prior snapshot. Scoring is host-shaped: TypeScript uses its
metric, Python/Java/Go can accept a metric callback, and all generated ports can
use task `score`/`scores` values plus the agent evaluation result.

## Playbook vs optimize()

- `playbook(...)` — accumulate reusable, evolving task knowledge; the only path that also learns online via `update(...)`.
- `optimize(...)` / `agent.optimize(...)` — tune instruction text and few-shot demos offline to a best/Pareto result.
- They are complementary; a project can use both.

## Troubleshooting

- "Cannot convert undefined or null to object" from `update()` → you passed input fields at the top level; wrap them in `example: { ... }`.
- Empty playbook after `evolve()` → the model already scored well, so nothing was curated; use harder/ambiguous examples or a weaker `studentAI` to surface lessons.
- Playbook not affecting an agent's behavior → ensure `apply` is not `false` and you used `agent.playbook(...)` (not a bare `playbook()` on an internal program).
- Playbook never grows with an `isExpensive` teacher → the reflector and curator calls are rejected before any request is sent, and only `verbose: true` logs it. Set `teacherOptions: { useExpensiveModel: 'yes' }`. In `apb.evolve(...)`, a rejected miner call is reported as a `miner failed` message through `onProgress` (and the console with `verbose: true`).

## See Also

- `ax-gepa` - `optimize(...)` and `AxGEPA` for instruction/demo tuning.
- `ax-agent-context` - choosing between contextMap, contextPolicy, `agent.playbook(...)`, and recall.
- `ax-agent-optimize` - `agent.optimize(...)` GEPA tuning for agents.