How we use coding agents at Layercode
Claude Code and Codex has been speeding up dev at Layercode, but also led us down time wasting rabbit holes.
Here’s what we’ve learned so far:
1. As of GPT5, Codex is succeeding in tasks more often than Claude.
I like to use high reasoning and trade time for more confidence that the code is correct.
“npm i @openai/codex@native” will install the new rust version being built.
2. Every task we write now has an AI Level:
- AI Assist code manually but using Cursor Tab and sidebar chat to help
- AI First means starting with an AI agent to write the first draft, or even the complete task if it can.
- Human Only means the task is expected to be complex and should be done completely by hand. This includes disabling Cursor Tab in some cases, as we have found it can introduce sneaky bugs in code it doesn’t fully understand.
AI First, tasks are now written as prompts that can be copy pasted into Claude or Codex.
3. You can often do 2-3 parallel AI First tasks at a time, reviewing one whilst the other is responding.
There are neat tools like claude-squad and conductor which spin up git worktrees for each. But to run your app in there you need to setup env, db etc
If, as a human you’re only going to do max 3 parallel tasks, the simple solution is just checkout your repo in 3 folders, setup each env and just run Claude or Codex in each.
I'm keen to hear from others about what what has and hasn't been working.
Replies
The single most important thing that has helped improve my results with AI coding agents (primarily Claude Code) is establishing context before asking for code changes. So instead of "Change the code so that it does X after Y", I will start with "This code does Y by querying the database for blah. It then uses the result to display a page that says foo. Do you understand? Please ask me any questions that would make this more clear".
Layercode
Amen @audiodude, context is absolutely king when it comes to ensuring LLMs remain focused and useful!
Oh this is a really cool write-up, @dctanner. Absolutely love how you leverage both @Claude Code and @Codex by OpenAI. It's definitely inspiring.
Curious to learn more about your prompting strategies.
AI tools speed up development, but can introduce subtle bugs. What strategies do others use to improve AI-generated code accuracy?