My Claude Code tips and links
Note: I wrote these tips before reading the Claude Code Best practices document. It is good. Maybe read that first.
I started using Claude Code ~July 2026 when the company switched from Copilot to Claude. Here are some of my tips. Please share yours as well. These are more on the end of the "use AI to help you code" spectrum, as opposed to "loop, graph, and let AI do all the things", in part due to the company having a limited number of tokens to use. I mostly stopped using subagents when Copilot switched to usage-based billing in June 2026 since they are no longer free and I have to preserve my precious tokens. Also these notes come from the context of a frontend developer.
Claude Code tips
- Use Pi instead. https://pi.dev/. ;)
- Stay in the "Smart Zone" if it exists. Matt Pocock says the "smart zone" ends at about 150K, so I try to stay below that. See also his AI Engineer talk.
- Trim the system prompt. https://x.com/mattpocockuk/status/2074464823232888987. See also Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k. Note: Claude might have cut their system prompt some since this discussion.
- Write long prompts. Maybe use voice. Trim the system prompt, but make your prompt longer. https://x.com/thdxr/status/2054046560904745337.
- Set output style to Concise. Make your prompt longer, but make Claude's output shorter. https://x.com/ClaudeDevs/status/2090245922685063634.
- Customize the status line. You can tell Claude Code to customize the status line. I added total cost and tokens used. The status line is set in
~/.claude/settings.json. Example:{"statusLine": {"type": "command", "command": "bash ~/.claude/statusline-command.sh"}}. - Use
anti-slopfor TypeScript. https://x.com/dillon_mulroy/status/2087537790061850984. - Tell the agent to ask you questions. Maybe use Matt Pocock's skills. Matt Pocock's
/grill-meor/grill-with-docswas good and eye opening. I don't always want a long grilling session, so I've been using this prompt recently: "**If you are not clear what I mean, or if there are multiple options that require my opinion, ASK ME QUESTIONS BEFORE PROCEEDING. DO NOT proceed unless you have clear direction. DO NOT ASSUME YOU KNOW WHAT I WANT.**" and have been happy with the results. Matt Pocock's skills: https://x.com/mattpocockuk/status/2075218406266036236. - Ask the agent questions. Tell the agent to ask you questions and ask it questions. Have a discussion. (This is much more incentivized when you are billed per token vs. per request.) This is largely influenced by Mario Zechner but I can't find a good link.
- Plan, but maybe don't use Plan Mode. https://x.com/thorstenball/status/2081641642306437222. Or maybe don't plan.
- Maybe don't use memory. https://x.com/pidotdev/status/2088951405155426757. https://x.com/badlogicgames/status/2076638642529800396. https://x.com/Vjeux/status/2091380216660922663.
- Maybe hand write AGENTS.md and skills. https://x.com/theo/status/2082009220631953782.
- Maybe use Opus 4.8 instead of Opus 5. I've been mainly using Opus 4.8 as it fits my style of work better than Opus 5. Also I may have been influenced by some people.
- Find out what AI is good at. https://x.com/davis7/status/2087357329054605597. I've found AI to be good at debugging, research, and teaching. It is good at coding in small contexts, but is worse in larger contexts like architecture. It's also good for code review and for working toward a goal. Examples: Autoresearch isn't just for training models and Rewriting Bun in Rust.
- Maybe use Ghostty + Herdr. Ghostty is very fast and Cmd+Shift+Enter is "genuinely useful". Herdr replaces tmux and plays a nice bell when my agents are done. This space is changing quickly now. I switched from iTerm + tmux → Ghostty + tmux → cmux + tmux → Ghostty + herdr in less than a year. There are other things I haven't tried or don't know about. And Mitchell Hashimoto is starting a new thing so I'll probably switch to that when available.
TODO
- Use
claude --bare. This looks like a good step to make Claude Code closer to Pi. I tried runningclaude --bareand/contextshowed 0 tokens used, which I didn't know what possible with Claude Code. Normally I am starting at 11.3k tokens in a fresh session.
Questions
ast-grep. Pi + gpt-5.6-sol happily usedast-grepwith thisast-grepskill I found a while ago, but Claude Code doesn't like to use it. Anyone have success using it with Claude Code?
