OpenAI SDK Patterns

Quick Guide: Use the official openai npm package (v6+) to interact with OpenAI's API directly. Use client.responses.create() (Responses API) for new projects with built-in tools and server-side state, or client.chat.completions.create() (Chat Completions) for stateless chat flows. Use zodResponseFormat and client.chat.completions.parse() for structured outputs. Use .stream() or stream: true for streaming. Supports GPT-5.x family, GPT-4o, o4-mini, embeddings, vision, audio, and batch processing.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use the Responses API (client.responses.create()) for new projects -- it provides better performance, built-in tools, and server-side conversation state)

(You MUST use zodResponseFormat() from openai/helpers/zod for structured outputs -- do NOT manually construct JSON schemas)

(You MUST handle errors using OpenAI.APIError and its subclasses -- never use bare catch blocks without error type checking)

(You MUST configure appropriate retries and timeouts for production use -- the SDK retries 2 times by default on 429/5xx errors)

(You MUST never hardcode API keys -- always use environment variables via process.env.OPENAI_API_KEY)

</critical_requirements>

---

Auto-detection: OpenAI, openai, client.chat.completions, client.responses.create, client.responses.parse, client.embeddings, client.audio, zodResponseFormat, zodTextFormat, zodFunction, zodResponsesFunction, runTools, GPT-5, GPT-4o, o4-mini, gpt-5-mini, text-embedding-3, whisper, tts, OPENAI_API_KEY, toFile

When to use:

• Building applications that call OpenAI models directly (GPT-5.x, GPT-4o, o4-mini, etc.)
• Implementing chat completions with streaming responses
• Using the Responses API for agentic workflows with built-in tools (web search, file search, code interpreter)
• Extracting structured data from LLM responses with Zod schema validation
• Implementing function calling / tool use with the Chat Completions or Responses API
• Creating embeddings for RAG pipelines or semantic search
• Processing images with vision models or audio with Whisper/TTS
• Running batch jobs for high-volume, cost-efficient processing

Key patterns covered:

• Client initialization and configuration (retries, timeouts, proxies)
• Chat Completions API (messages, streaming, function calling)
• Responses API (input, instructions, built-in tools, server-side state)
• Structured outputs with zodResponseFormat and client.chat.completions.parse()
• Streaming with for await...of, .stream() helper, and event handling
• Embeddings API (text-embedding-3-small, text-embedding-3-large)
• Vision (image URLs, base64), Audio (Whisper transcription, TTS), Batch API
• Error handling, retries, timeouts, and production best practices

When NOT to use:

• Multi-provider applications where you need to switch between OpenAI, Anthropic, Google, etc. -- use a unified provider SDK instead
• React-specific chat UI hooks (useChat, useCompletion) -- use a framework-integrated AI SDK
• When you need a higher-level abstraction over multiple LLM providers

---

Examples Index

• Core: Setup & Configuration -- Client init, production config, Azure, error handling, request overrides
• Chat Completions -- Basic chat, multi-turn, token tracking, output length control
• Streaming -- stream: true, .stream() helper, Responses API streaming, abort
• Tool/Function Calling -- Manual tools, zodFunction, runTools automation, Responses API tools
• Structured Output -- zodResponseFormat, zodTextFormat, refusal handling
• Embeddings, Vision & Audio -- Semantic search, image analysis, transcription, TTS, batch processing
• Quick API Reference -- Model IDs, method signatures, error types, streaming events

---

<philosophy>

Philosophy

The official OpenAI SDK provides direct, low-level access to OpenAI's full API surface. It is the thinnest possible wrapper over the REST API, auto-generated from OpenAI's OpenAPI specification using Stainless.

Core principles:

1. Direct API access -- No abstractions or provider layers. You get the exact API that OpenAI documents, with full TypeScript types. Every API feature is available immediately when OpenAI releases it.
2. Two API paradigms -- The Responses API (client.responses.create()) is the newer, recommended API with built-in tools and server-side state. The Chat Completions API (client.chat.completions.create()) remains fully supported for stateless chat flows.
3. Built-in resilience -- The SDK handles retries (2 by default on 429/5xx), timeouts (10 min default), and auto-pagination out of the box.
4. Streaming as a first-class pattern -- Use stream: true for SSE-based streaming, .stream() helper for event-based consumption, or for await...of for simple iteration.
5. Type-safe structured outputs -- zodResponseFormat() and client.chat.completions.parse() convert Zod schemas to JSON Schema and parse responses, giving you validated, typed objects.

When to use the OpenAI SDK directly:

• You only use OpenAI models and want the simplest, most direct integration
• You need access to OpenAI-specific features (Responses API, Batch, Realtime)
• You want minimal dependencies and zero abstraction overhead
• You need the latest API features on day one

When NOT to use:

• You need to switch between providers (OpenAI, Anthropic, Google) -- use a unified provider SDK
• You want React-specific chat UI hooks -- use a framework-integrated AI SDK
• You want a higher-level agent framework -- consider OpenAI Agents SDK (@openai/agents)

</philosophy>

---

<patterns>

Core Patterns

Pattern 1: Client Setup

Initialize the OpenAI client. It auto-reads OPENAI_API_KEY from the environment.

// lib/openai.ts -- basic setup
import OpenAI from "openai";
const client = new OpenAI();
export { client };

// lib/openai.ts -- production configuration
const TIMEOUT_MS = 30_000;
const MAX_RETRIES = 3;
const client = new OpenAI({ timeout: TIMEOUT_MS, maxRetries: MAX_RETRIES });

Why good: Minimal setup, env var auto-detected, named constants for production settings

See: examples/core.md for Azure OpenAI, per-request overrides, error handling patterns

---

Pattern 2: Chat Completions API

Stateless text generation. You manage conversation history.

const completion = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [
    { role: "developer", content: "You are a helpful coding assistant." },
    { role: "user", content: "Explain TypeScript generics." },
  ],
});
console.log(completion.choices[0].message.content);

Why good: Clear message roles, developer message for system instructions, direct content access

// BAD: No developer message, no error handling
const res = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "do something" }],
});

Why bad: No system instruction means unpredictable behavior, vague prompt

See: examples/chat.md for multi-turn, token tracking, output length control

---

Pattern 3: Responses API (Recommended for New Projects)

Newer API with built-in tools, server-side state, and better performance with reasoning models.

const response = await client.responses.create({
  model: "gpt-4o",
  instructions: "You are a coding assistant.",
  input: "What are TypeScript generics?",
});
console.log(response.output_text);

Why good: Clean separation of instructions and input, output_text helper, simpler than messages array

// BAD: Using Chat Completions parameters with Responses API
const response = await client.responses.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello" }], // WRONG: use 'input'
});

Why bad: Responses API uses input and instructions, not messages

Built-in Tools

Web search ({ type: "web_search_preview" }), file search ({ type: "file_search" }), code interpreter ({ type: "code_interpreter" }). Chain conversations with previous_response_id and store: true.

See: examples/tools.md for Responses API function calling with tool outputs

---

Pattern 4: Streaming

Use streaming for user-facing responses.

// Chat Completions -- stream: true with for-await
const stream = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Explain async/await." }],
  stream: true,
});
for await (const chunk of stream) {
  const content = chunk.choices[0]?.delta?.content;
  if (content) process.stdout.write(content);
}

// Event-based with .stream() helper
const stream = client.chat.completions.stream({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Tell me a story." }],
});
stream.on("content", (delta) => process.stdout.write(delta));
const finalContent = await stream.finalContent();

Why good: Progressive output for better UX, event-based API for granular control

// BAD: Not consuming the stream
const stream = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello" }],
  stream: true,
});
// Stream never consumed -- tokens are lost

Why bad: Stream must be consumed via iteration or event handlers, otherwise tokens are lost

See: examples/streaming.md for Responses API streaming, abort, stream methods

---

Pattern 5: Structured Outputs with Zod

Use zodResponseFormat() and .parse() for type-safe structured responses.

import { zodResponseFormat } from "openai/helpers/zod";
import { z } from "zod";

const CalendarEvent = z.object({
  name: z.string(),
  date: z.string(),
  participants: z.array(z.string()),
});

const completion = await client.chat.completions.parse({
  model: "gpt-4o",
  messages: [
    { role: "developer", content: "Extract event details." },
    { role: "user", content: "Alice and Bob meet next Tuesday for lunch." },
  ],
  response_format: zodResponseFormat(CalendarEvent, "calendar_event"),
});

const event = completion.choices[0].message.parsed; // Fully typed

Why good: Auto-converts schema, validates output, fully typed result, handles refusals

See: examples/structured-output.md for Responses API (zodTextFormat), refusal handling, complex schemas

---

Pattern 6: Function Calling / Tool Use

Define functions the model can call. Use zodFunction() for type-safe definitions.

import { zodFunction } from "openai/helpers/zod";
import { z } from "zod";

const GetWeatherParams = z.object({
  location: z.string().describe("City name"),
  unit: z.enum(["celsius", "fahrenheit"]).default("celsius"),
});

const completion = await client.chat.completions.parse({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Weather in Paris?" }],
  tools: [zodFunction({ name: "get_weather", parameters: GetWeatherParams })],
});

const toolCall = completion.choices[0].message.tool_calls?.[0];
if (toolCall?.type === "function") {
  console.log(toolCall.function.parsed_arguments); // Typed from Zod
}

Why good: zodFunction provides type-safe argument parsing, .describe() guides the model

Use runTools() for automated tool execution loops that handle the call-respond cycle automatically.

See: examples/tools.md for runTools, manual tool definitions, Responses API function calling

---

Pattern 7: Embeddings, Vision & Audio

• Embeddings: client.embeddings.create({ model: "text-embedding-3-small", input: [...] }) -- batch multiple inputs in one call
• Vision: Multi-part content array with { type: "image_url", image_url: { url } } for URL or base64 images
• Audio: client.audio.transcriptions.create() for speech-to-text, client.audio.speech.create() for TTS
• Files: client.files.create() with ReadStream, Buffer (via toFile), or fetch() Response
• Batch API: Upload JSONL, create batch with client.batches.create(), poll for completion at 50% cost

See: examples/embeddings-vision-audio.md for full examples with cosine similarity, base64 images, timestamps, TTS voice instructions, batch processing

---

Pattern 8: Error Handling

Always catch OpenAI.APIError and its subclasses. Re-throw unexpected errors.

try {
  const completion = await client.chat.completions.create({
    model: "gpt-4o",
    messages: [{ role: "user", content: "Hello" }],
  });
} catch (error) {
  if (error instanceof OpenAI.APIError) {
    console.error(
      `API Error [${error.status}]: ${error.message} (${error.request_id})`,
    );
    // Check subclasses: RateLimitError, AuthenticationError, BadRequestError, etc.
  } else {
    throw error; // Re-throw non-API errors
  }
}

Why good: Specific error types with status codes, request ID for debugging, re-throws unexpected errors

See: examples/core.md for full production error handling, stream errors, error type hierarchy

</patterns>

---

<performance>

Performance Optimization

Model Selection for Cost/Speed

General purpose             -> gpt-5.4 (most capable) or gpt-4o (proven, lower cost)
Cost-sensitive / high-vol   -> gpt-5-mini or gpt-5-nano (cheapest)
Complex reasoning           -> gpt-5.4 or o4-mini
Structured output           -> gpt-5.4 or gpt-4o (best schema adherence)
Embeddings                  -> text-embedding-3-small (cheapest) or text-embedding-3-large (highest quality)
Transcription               -> whisper-1 or gpt-4o-transcribe (higher accuracy)
TTS                         -> tts-1 (fast) or tts-1-hd (quality) or gpt-4o-mini-tts (voice control)
Batch processing            -> gpt-5-mini at 50% batch discount

Key Optimization Patterns

• Track token usage via completion.usage for cost visibility
• Check finish_reason === "length" to detect truncated output
• Use temperature: 0 for deterministic output (enables caching)
• Use AbortController to cancel long-running requests
• Use Batch API for high-volume jobs at 50% cost reduction

</performance>

---

<decision_framework>

Decision Framework

Which API to Use

Building a new application?
+-- YES -> Need built-in tools (web search, file search, code interpreter)?
|   +-- YES -> Use Responses API (client.responses.create())
|   +-- NO -> Need server-side conversation state?
|       +-- YES -> Use Responses API with store: true
|       +-- NO -> Either API works, prefer Responses for new code
+-- Existing Chat Completions code?
    +-- Working fine? -> Keep using Chat Completions (fully supported)
    +-- Need new features? -> Consider migrating to Responses API

Which Model to Choose

What is your task?
+-- General text generation -> gpt-5.4 (most capable) or gpt-4o (lower cost)
+-- Fast + cheap simple tasks -> gpt-5-mini or gpt-5-nano
+-- Complex reasoning / math -> gpt-5.4 or o4-mini
+-- Structured output -> gpt-5.4 or gpt-4o (best schema adherence)
+-- Vision (images) -> gpt-5.4 or gpt-4o
+-- Embeddings -> text-embedding-3-small (default) or text-embedding-3-large
+-- Transcription -> whisper-1 or gpt-4o-transcribe
+-- Text-to-speech -> tts-1 (fast) or gpt-4o-mini-tts (voice instructions)
+-- Batch processing -> gpt-5-mini (cheapest at 50% batch discount)

Streaming vs Non-Streaming

Is the response user-facing?
+-- YES -> Use streaming (stream: true or .stream())
|   +-- Need event-level control? -> .stream() with event handlers
|   +-- Simple text output? -> stream: true with for await
+-- NO -> Use non-streaming
    +-- Background processing -> client.chat.completions.create()
    +-- Structured output -> client.chat.completions.parse()
    +-- High volume -> Batch API

When to Use This SDK vs a Provider-Agnostic SDK

Do you need multiple LLM providers (OpenAI + others)?
+-- YES -> Not this skill's scope -- use a unified provider SDK
+-- NO -> Do you need OpenAI-specific features?
    +-- YES -> Use OpenAI SDK directly
    |   Examples: Responses API, Batch API,
    |   Realtime API, built-in web search/file search
    +-- NO -> OpenAI SDK is simplest for OpenAI-only use

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Hardcoding API keys instead of using environment variables (security breach risk)
• Using bare catch blocks without checking OpenAI.APIError (hides API errors)
• Not consuming streams returned by stream: true (tokens are silently lost)
• Using JSON.parse() on completion content without zodResponseFormat (fragile, no validation)
• Sending full conversation history every request when Responses API's previous_response_id could manage state

Medium Priority Issues:

• Not setting maxRetries / timeout for production deployments (10 min default timeout may be too long)
• Missing developer role message (no system instruction = unpredictable output style)
• Using deprecated system role instead of developer role in Chat Completions
• Not checking finish_reason for 'length' truncation
• Ignoring usage data (no cost visibility)

Common Mistakes:

• Confusing Responses API (client.responses.create()) with Chat Completions (client.chat.completions.create()) parameters -- they use different shapes
• Using messages parameter with Responses API (it uses input and instructions)
• Using response_format with models that don't support structured outputs (need gpt-4o or later)
• Using max_tokens with reasoning models (o4-mini, gpt-5.x) -- use max_completion_tokens instead
• Not handling the case where completion.choices[0].message.tool_calls is undefined
• Forgetting that runTools() defaults to max 10 completions -- set maxChatCompletions explicitly

Gotchas & Edge Cases:

• The SDK auto-retries on 429 (rate limit) and 5xx errors -- 2 retries by default. Disable with maxRetries: 0 if you handle retries yourself.
• stream: true returns raw SSE chunks. Use .stream() helper for a nicer event-based API.
• client.chat.completions.parse() throws LengthFinishReasonError if finish_reason is 'length' and ContentFilterFinishReasonError if 'content_filter'.
• Embedding responses return Array<number> (the SDK requests base64 by default and decodes via Float32 internally for performance). No conversion needed -- you get a plain number array.
• File uploads support ReadStream, File, fetch() Response, and toFile() helper -- use whichever matches your data source.
• The Responses API's store: true enables server-side state but also means OpenAI stores your conversations. Set store: false for sensitive data.
• developer role replaces system role in newer models (gpt-4o and later).
• Batch API has a 24h completion window and 50,000 request limit per batch.
• Audio transcription has a 25 MB file size limit.
• Zod schemas with zodResponseFormat must use additionalProperties: false -- the SDK handles this automatically.
• zodTextFormat and zodResponseFormat are NOT compatible with Zod v4 -- use Zod v3.x until the SDK adds v4 support.
• The Assistants API is deprecated (sunset August 2026) -- use the Responses API for new code.

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use the Responses API (client.responses.create()) for new projects -- it provides better performance, built-in tools, and server-side conversation state)

(You MUST use zodResponseFormat() from openai/helpers/zod for structured outputs -- do NOT manually construct JSON schemas)

(You MUST handle errors using OpenAI.APIError and its subclasses -- never use bare catch blocks without error type checking)

(You MUST configure appropriate retries and timeouts for production use -- the SDK retries 2 times by default on 429/5xx errors)

(You MUST never hardcode API keys -- always use environment variables via process.env.OPENAI_API_KEY)

Failure to follow these rules will produce insecure, unreliable, or poorly-typed AI integrations.

</critical_reminders>