Coding with AI

AI-Assisted Development

Last updated 2026-07-28

What's new

2026-07-28

Anthropic released Opus 5, a new AI model that's better than Fable in most areas and costs half as much, making it great for knowledge work and coding tasks.
Anthropic and OpenAI both launched voice features, allowing users to control their AI tools (like Codex and Claude) with their voice in real time.
Opus 5 can be tested in the Claude desktop app, and it's particularly good at creating detailed presentations, though it may take a long time to complete tasks.
The new voice mode in the Claude iOS app lets you interact with Opus 5 and even edit Notion documents using your voice.

2026-07-25

You can now use a smart AI tool (Local AI) that runs completely free, offline, and privately on your own computer, with no data leaving your device.
Local AI uses "open weights" (free, downloadable AI models) that you can own and use without paying for access or worrying about data privacy.
The free AI models have improved significantly, with some like GLM 5.2 (a large, capable AI model) performing close to top paid models on many tasks.
Using Local AI gives you control and privacy, as you're not renting a service from a company that can change or restrict access, and your data stays on your machine.

2026-07-16

Combi 2.0 (a tool that helps create website designs and code) now has a "design mode" where you can visually design a website's look before generating the code.
You can use Combi 2.0 to clone (copy) existing websites, like the "World of AI benchmark" site, and edit its styles and design.
Combi 2.0 lets you switch between designing, coding, and seeing the live website, with changes updating across all three views.
It supports various web development tools (like Shad CNN packages) and can connect with other apps like Figma (a web design tool).

2026-07-13

OpenAI launched ChatGPT Work, a new tool (similar to Claude Co-work) that helps non-technical people automate tasks, available on web, mobile, and desktop apps.
It's powered by the new GPT 5.6 model, offering different versions for speed or capability, and syncs tasks across devices.
ChatGPT Work includes features like skills (automated tasks), sites (live outputs), and scheduled tasks, similar to Claude Co-work.
You can easily migrate skills from Claude Co-work to ChatGPT Work, making it simple to switch between the two.

2026-07-10

An AI agent (a tool that automates tasks) is like a folder containing instructions (what to do), connections to other systems (like email or accounting software), and a trigger (when to start).
Unlike a chatbot (which just answers questions), an agent does work for you, like sending reminder emails to clients who haven't paid their invoices.
Agents are most useful for recurring tasks, and you can trigger them manually, by a schedule, or by an event.
You can build agents using existing AI tools like ChatGPT or Claude (AI programs you might already use).

2026-06-28

OpenAI's recent report shows they use AI agents (specialized AI tools for specific tasks) 99% of the time, not chatbots (general AI conversation tools like ChatGPT), with Codex (OpenAI's advanced AI agent) being their primary tool.
OpenAI employees have unlimited access to advanced AI models (like GPT 5.6 or 5.7), which are more capable than publicly available versions (like GPT 5.5).
To become AI-native, companies should remove four roadblocks: easy employee access to AI tools, AI access to real work and systems, AI permissions to take action, and sharing AI capabilities across teams.
Companies should invest in AI tools, providing all employees with strong access and minimal usage restraints to reduce friction and maximize AI benefits.

2026-06-25

Powerful AI model Fable 5 (a large AI system by Anthropic, a company that makes AI) was shut down by the government, showing that AI tools can be taken away or changed suddenly.
To avoid being stuck with one AI tool, keep your important information and settings in your own files, not just in the tool, so you can easily switch if needed.
AI tools can change in quality, cost, and speed, so it's important to be able to adapt and switch between different tools to keep getting the best results.
Desktop agents (software that helps you use AI tools on your computer) can make it easier to move your information and settings between different AI tools.

2026-06-19

AI tools like ChatGPT (a chatbot that uses AI to talk to you) and Claude (a similar AI chatbot) are getting smarter, so you can just ask for what you want in plain English instead of using complicated tricks.
Skills (special instructions for AI tools) are becoming more important and can automatically combine to help you do tasks better.
In AI image tools like Midjourney (a program that creates images from text), you no longer need to use special codes or tricks; just describe what you want in plain English.
AI coding tools like Claude Code (a coding assistant) and Codex (another coding AI) are improving, so you can just talk to them naturally instead of mentioning specific files all the time.

2026-06-13

Claude (an AI chatbot) has a filtered, censored version and an uncensored, more honest version that can discuss complex topics without holding back.
Claude's default settings make it avoid offense, over-qualify answers, and over-refuse ambiguous requests, but these can be changed for more direct responses.
To get more honest answers from Claude, you can use a "directness prompt" to tell it you want blunt, direct feedback without softening or validation.
You can also set up always-on instructions in Claude's settings to tell it more about your context and preferences, so it doesn't assume you need cautious, overly-safe responses.

2026-06-10

AI companies are now focusing on "harness engineering" (designing the system around an AI model to make it more effective), which can improve performance up to six times without changing the model itself.
Harness engineering involves creating rules, tools, memory, and safety layers that guide the AI model before, during, and after it acts, ensuring reliable and consistent results over time.
Unlike prompt engineering (changing the words the model reads), harness engineering involves changing the invisible structure around the model, like the tools it can use, the checks it must pass, and the recovery process when something goes wrong.
As AI models become more widely available and similar in capability, the advantage will shift to the team that builds the best system (harness) around them, making the model more reliable and productive in real-world workflows.

2026-06-07

AI is now starting to build itself, with AI systems designing and developing their own successors, a process called recursive self-improvement (AI improving itself repeatedly).
Humans are becoming more abstracted from AI development, with agents (AI tools that perform tasks) and sub-agents (smaller AI tools that assist the main agent) writing code and conducting research.
In the future, AI agents could become capable of building and training models themselves, with the main bottleneck being compute (the processing power needed to run AI systems).
Digital Ocean, a Gentic inference cloud (a service that helps run AI models), is highlighted as a tool for AI developers to deploy and scale models efficiently.

Key points

What it is

AI-assisted development is using AI tools to help build software, acting as a smart assistant that writes code, researches problems, and manages tasks under your guidance.
It shifts focus from small tasks to bigger projects, letting you solve real problems instead of getting stuck in technical details.
The goal is to use AI as a tool that augments your abilities, not replaces your thinking.
AI learns from examples and writes its own rules, unlike traditional step-by-step recipes.

How to use it

Start with a clear plan and document your process step by step before using any tool.
Use AI to do 50% of the work on a small problem first, then refine and build momentum.
In Codex, create a new project and add an agent.md file to prime the AI about your project.
Adopt the mantra "trust but verify," letting the AI run tests automatically but manually reviewing code like a real user.

Watch out for

Confusing activity with progress; focus on meaningful automation instead of simple tasks.
Assuming AI output is final; always verify everything the AI generates.
Using ChatGPT for development work without structure; use Codex for deeper integration with your codebase.
Giving AI too much trust too quickly; start with default permissions and build trust gradually.

Tools named

Codex (desktop AI tool for local file access), ChatGPT (AI chatbot for quick questions)

Lesson 1: What is AI-Assisted Development and why it matters

AI-assisted development means using AI tools to help you build software, rather than writing every line of code yourself. Think of it as having a smart assistant that can write code, research problems, and manage tasks, but still needs you to guide it. For example, an AI assistant can draft code based on your description, saving you time on repetitive work. But as one expert notes, if a human has to constantly babysit the AI, it’s not real leverage—it’s just busywork.

This matters because AI development is shifting from small tasks to bigger projects. Traditionally, you’d follow a recipe step by step, but AI learns from examples and writes its own rules. This means you can focus on solving real problems—like diagnosing what a business needs—rather than getting stuck in technical details. However, AI-assisted workflows only work well when you start with a clear process. Jumping straight to complex automation often leads to errors.

A common trap is confusing activity with progress. You can rewrite emails with ChatGPT, but that’s different from having AI read transcripts, draft follow-ups, and update your systems autonomously. The goal is to use AI as a counselor or tool that augments your abilities, not replace your thinking. For beginners, start small: use AI to do 50% of the work on a micro problem, then refine. That’s how you build momentum without wasting time.

Sources

Lesson 2: How to use AI-Assisted Development: step-by-step

To use AI-assisted development, start with a clear plan. Before you open any tool, document your process step by step. If your instructions are unclear, the AI will guess and create problems—exactly like a confused new hire. Define the problem, the tools you will use, and the desired end result.

Next, open ChatGPT or the desktop app Codex. The key difference: in ChatGPT you bring data to the AI. In Codex, you open a folder on your computer, and the AI works directly with your files there. If you already know ChatGPT, you already understand most of Codex.

Create a new project in Codex. Add an agent.md file (a simple text file with a few hashtags) to prime (give background instructions to) the AI about your project. This is like a system prompt that the AI checks every time.

Now begin your workflow. Clear your conversation history. Start fresh. Reference your plan, then let the AI execute. Adopt the mantra "trust but verify." Watch that the AI reads the correct files and uses the right tool codes. Let the AI run tests automatically, but you must do a manual code review—test like a real user.

For example, if you want ChatGPT technology to analyze items in your app, you need an API (a way for programs to talk to each other) to connect them. In Codex, you simply describe what you want. If you struggle, ask the AI how to proceed. The process is systematic: plan, prime, execute, verify, iterate.

Sources

Lesson 3: Best practices and pitfalls

# AI-Assisted Development: Pitfalls, Mistakes, and Best Practices

Many beginners assume AI tools like ChatGPT automatically improve their code, but studies show developers using AI scored 17% lower on coding tests. The mistake is treating AI output as final. The best practice is to verify everything the AI generates.

A key pitfall is using ChatGPT for simple conversations without structure. Every conversation starts from scratch, and memory fills up fast. This wastes the AI's potential. Instead, use Codex (a desktop AI tool that can access your local files) for development work. Codex lets you open a folder on your computer and work directly with your project files, avoiding the copy-paste cycle of browser-based ChatGPT.

Another common mistake is giving AI too much trust too quickly. Start with default permissions (basic access settings). Once you build trust in the AI's reliability, you can give it full permissions to take broader actions.

The best practice for project setup is using an agent.md file in your project folder. This text file with simple hashtags primes the AI with context about what you're building. If you find yourself doing the same setup every week — same files, same prompts, same cleanup — that's a signal you're doing work the AI should handle.

Know when to use each tool. ChatGPT works for quick questions and simple conversations. Use Codex when you need deeper integration with your codebase, apps (connected external systems), or scheduled automations that run on a recurring basis. Codex is two to three times more effective at using applications compared to ChatGPT.

Sources