Developer Guides
Guides written for developers, covering implementation details and trade-offs.
Articles in Developer Guides (page 1 of 2)
- Building Your First AI Agent with the GPT-6 Astra API: A Step-by-Step Developer Guide — A hands-on walkthrough of building a tool-calling research agent with GPT-6 Astra, covering tool definitions, the reasoning loop, structured output, error handling, and where human checkpoints belong.
- GPT-6 Astra API: A Complete Developer's Guide to Pricing, Context, and Integration — A practical breakdown of the GPT-6 Astra API's context window, pricing, and tool-calling behavior for developers deciding whether to migrate an existing GPT-5.6 Sol integration.
- Gemini API Long-Running Operations: The Async Polling Pattern Explained — File uploads, Batch API jobs, and File Search indexing all use the same underlying pattern: kick off an operation, poll until it is done. This guide unifies what earlier posts each covered individually into one reusable pattern.
- Using the Gemini API with the Vercel AI SDK — The Vercel AI SDK gives Gemini a framework-native interface for Next.js and React apps -- generateText, streamText, generateObject, and tool calling, all working with Gemini models through one provider package.
- Migrating from OpenAI to the Gemini API: A Practical Guide — Google ships a real OpenAI-compatible endpoint for the Gemini API, so an existing OpenAI integration can point at Gemini by changing a base URL. Here is what works out of the box, what needs a real code change, and what to check before you ship.
- Evaluating and Comparing Gemini Models: A Practical Framework — With a dozen-plus Gemini models spanning Flash, Pro, and Lite variants, picking one by reading spec sheets is not enough. Here is how to actually compare them -- AI Studio Compare Mode for quick checks, and a DIY harness for production decisions.
- Building Multi-Tool Agents with the Gemini API — Function calling, code execution, grounding, and File Search each work well alone. An agent is what happens when a model decides, turn by turn, which of them to reach for. Here is how to build one, manually and with Google ADK.
- Using the Gemini API with LangChain: A Practical Integration Guide — LangChain is one of the most common ways developers wire Gemini into larger applications -- chains, agents, retrievers. Here is how langchain-google-genai actually works, and how it relates to the official google-genai SDK.
- Gemini Embedding Task Types and Output Dimensionality Explained — The same text embedded for search versus for classification should not produce identical vectors. task_type and output_dimensionality are how the Gemini Embedding API lets you tune for the job -- most integrations never touch either.
- Calling the Gemini API via REST Without an SDK — Not every language has an official Gemini SDK. The REST API underneath works the same everywhere -- here is the exact endpoint, headers, and request body, with working curl and a language-agnostic pattern for any HTTP client.
- Understanding Gemini API finish_reason and Response Metadata — A Gemini API call can return HTTP 200 and still hand you a response that is not what you asked for -- truncated, blocked, or cut off mid-thought. finish_reason is how you find out, and most integrations never check it.
- Gemini API Generation Config Deep Dive: Temperature, Top-P, Top-K, and More — Every generateContent call accepts a GenerationConfig object, and most developers only ever touch temperature. Here is what each parameter actually controls, when to change it, and when Google explicitly says not to.
- Prompt Engineering Best Practices for the Gemini API — The same Gemini model gives noticeably different results depending on how you ask. Here is what Google actually documents about structuring prompts, writing system instructions, and using few-shot examples -- not general folklore.
- Gemini Video Understanding: Analyzing Video Content with Timestamp References — Gemini can watch a video, not just generate one -- sampling it at 1 frame per second, reasoning over audio and visuals together, and answering questions tied to a specific timestamp. Here is how native video analysis works.
- Gemini API Token Counting and Context Window Management — A request that silently exceeds a model context window fails at the worst possible time -- in production, on your longest and most important input. Here is how to count tokens ahead of time and manage the window deliberately.
- Streaming Responses with the Gemini API: Building Real-Time Text UIs — Waiting for a full Gemini response before showing anything makes even a fast model feel slow. Streaming sends the answer token by token instead. Here is the event protocol and working Python and JavaScript.
- Git Workflows With Claude Code: Letting an Agent Commit Without Losing Control — The rules that make git workflows with Claude Code trustworthy aren't complicated: never rewrite history you don't own, and let the diff speak for the what so the message can speak for the why.
- Using Claude Code for Code Review: What It Catches That a Linter Doesn't — A linter checks the code you wrote against a rule. Using Claude Code for code review checks the code you wrote against what you meant — a fundamentally different kind of catch.
- Claude Code Permission Modes: Balancing Speed and Safety — The right Claude Code permission mode isn't about trusting the model more or less — it's about how expensive a wrong action would be in this specific repo.
- Custom Slash Commands in Claude Code: Turning Repeated Prompts Into Workflows — If you've typed a variation of the same paragraph into Claude Code three times, it should be a slash command — a markdown file, not a mental habit only you remember.
- Plan Mode in Claude Code: When It's Worth the Extra Step — Plan Mode trades speed for a checkpoint before code changes happen. That trade is worth it exactly when the ambiguity is in the approach, not the syntax.
- Claude Code Hooks: Automating Guardrails Without Slowing Yourself Down — A PreToolUse hook that blocks a dangerous command is worth the friction. A hook that fires on every single edit is not. Where that line actually sits after using both.
- CLAUDE.md: What Actually Belongs in Your Project Memory File — CLAUDE.md gets read fresh every session, which means every line in it has a cost. What survives the edit after your third rewrite is usually the only part worth keeping.
- Claude Code Context Window: Managing Compaction Without Losing the Thread — Auto-compaction summarizes your conversation before it overflows — but what it keeps and what it silently drops changes how you should structure a long Claude Code session.