If you're a dev or interested in keeping up with AI news, tools, and trends, @ConnorFitzDev and myself have been working on a weekly AI roundup newsletter geared towards busy devs wanting to stay ahead of the curve. Get caught up in 5min every Tuesday:

@METR_Evals Hopefully this has been helpful! If you have any other tips, feel free to share them, I'd love to hear them. If you're interested in reading the research that sparked this thread, you can find it here:

@METR_Evals Pick your battles. Not everything needs AI. For example, if creating a quick util function takes longer to write the prompt & address the issues than it would to just write it yourself, you should have written it yourself. Learn where it makes sense to use & where it doesn't.

@METR_Evals AI code generation quality typically falls off if you're trying to get 100% task completion with it (as complexity grows at least). I tend to aim for anything in the range of 70-80% completion and then I manually jump in to finish the rest. Find where the sweet spot is for you.

@METR_Evals On a similar note, if you're using Claude Code, don't forget to leverage the context / conversation clearing feature. Entering `/clear` can sometimes help with the output quality if you've been working within the conversation for too long on too many different, unrelated tasks.

@METR_Evals Don't try and do it all at once. Break down complex tasks into smaller, manageable chunks (this is also generally good engineering practice anyway). AI can quickly get overwhelmed if you're trying to do too much at once and the quality of the code output will suffer.

@METR_Evals Create checkpoints to easily go back in time. Tools like Cursor have this built in, but Claude Code doesn't quite have the same. Pro tip: use Git commits to create checkpoints for Claude Code so you can easily revert if needed rather than have the AI try & fix itself forever.

@METR_Evals If you're using Claude Code, keep your up to date with instructions around common mistakes you're seeing or patterns you want it to follow. This will save you a ton of time and ensure it's not producing the same frustrating mistakes over and over again.

@METR_Evals Always plan first before jumping into more complex tasks. For example, Claude Code allows you to Shift+Tab to enter Plan mode. Start there to ensure it has the correct context and understands what you want before it starts to generate any code. It'll save you a lot of time.

A new circulating @METR_Evals study found devs using AI tools like Claude Code were 19% slower in producing code. Ever been stuck in a loop of bad code generation? You’re not alone. Here are some tips I’ve found to make sure AI speeds you up instead of slowing you down: 🧵.

@AnthropicAI And that's about it! You can see there's a lot of personal preference involved in Claude Code workflows, so experiment and find what works for you. If you want to dive a bit deeper, you can read the full article from their engineering blog here:

@AnthropicAI You can have multiple instances of Claude Code running at once. This allows for things like having one instance write code while another verifies it (typically just run /clear). Another one is having multiple checkouts of your repo in different folders & start Claude in each one.

@AnthropicAI Take advantage of headless mode in Claude Code when you don't need it to be interactive, such as for CI, pre-commit hooks, etc. This is extremely helpful for issue triaging, linting, and so on.

@AnthropicAI Take advantage of checklists and scratchpads when things start to get complex. For example, when working through something like a code migration, you can have Claude Code create and use an MD file as a checklist and working scratchpad.

@AnthropicAI Skip auto-accept mode initially so you can see the reasoning and changes it's proposing. If you really want to use it, I recommend starting with plan mode. You can interrupt it (Esc) and even undo changes. You can also clear the context window to help it stay focused (/clear).

@AnthropicAI Much like any other LLM tool, be specific. The more specific you are with Claude Code, the better your results will be on the first attempt. In order to help with that, give Claude Code images, mention specific files, give it URLs, pass it data (via Copy/Paste or MCP) and so on.

@AnthropicAI There aren't specific workflows for Claude Code, so try out some of the common ones to see what works, such as:. 1. Explore → plan → code → commit.2. Write tests → commit → code → iterate → commit.3. Write code → screenshot result (via MCP server) → iterate. And so on.

@AnthropicAI Use custom slash commands to improve your repetitive workflows. Think of things like debugging, analyzing logs, and so on. Custom slash commands are more or less just MD prompt templates stored in the .claude/commands folder of your project.

@AnthropicAI Give it more power. Claude Code has access to your shell (+ inherits your bash env), so tell it about custom bash tools. It also acts as an MCP server & client, so it can connect to any number of MCP servers & their tools via project config, global config, or an .mcp.json file.

@AnthropicAI Customize your setup by creating a in your repo to preload project-specific context like test instructions, folder structure, repo etiquette, etc. Keep it updated & iterate on it. You can also run `/init` to have Claude automatically generate it for you.

There's a lot of hype around Claude Code (for good reason), but it often gets tricky to figure out what the current "meta" is. Start with best practices & iterate until you find the perfect workflow for you. Here's a quick overview of best practices @AnthropicAI talk about 👇🧵.