20 October 2026
خمس عادات من اللي بنى Claude Code: أوامر وبرومبتات جاهزةFive Habits from the Engineer Who Built Claude Code: Ready Commands and Prompts
Five habits from Boris Cherny, the engineer who built Claude Code, turned into commands and prompts you can use in your next session.
Who this guide is for
Anyone working in Claude Code, from week one to running agents every day
Key terms
- Claude Code
- Claude in your terminal. It reads and edits your files and runs commands itself.
- CLAUDE.md
- An instructions file in your project that Claude Code reads every session. What you put there sticks.
- skill
- A folder of instructions for one job. Claude loads it when needed, so you stop explaining it every time.
- context
- Everything in front of the model right now: prompt, tools, files and the conversation. The more there is, the more it drifts.
The steps
Step 1
Give it the task, skip the plan
Boris says the newer models don't need a planning step anymore. Instead of writing a long plan, state the task and the result you want, and let it work.
For a big change that touches many files, use /plan first and review it before it starts.
PromptDirect task[the task in one sentence], then run [the tests or command that proves it works] and tell me the result.
Tip: Always end the task with a way for it to check itself, like the tests.
Step 2
Make it write its mistake into CLAUDE.md
When Claude makes a mistake that keeps coming back, don't just fix it yourself. Tell it to write a rule in CLAUDE.md so the next session doesn't repeat it.
Type /memory to open the file and edit the rules yourself.
PromptLog the mistakeWhat just went wrong: [the mistake]. Add a short rule to CLAUDE.md so it doesn't happen again, and show me the rule you added.
Step 3
Turn a repeated job into a skill
If you type the same request every week, make it a skill. Boris says this is what lets Claude keep running without stopping to ask you.
PromptMake a skillMake me a skill called [job name] with the steps we just did in this session and the mistakes we avoided, so next time I can call it with one word.
Step 4
Talk to a loop, not an agent
Instead of checking back every few minutes, let a loop prompt for you. /loop runs the same prompt on a schedule while the session stays open.
Give an interval, then the prompt. With no interval, Claude picks the timing itself.
PromptFollow up loop/loop 5m check if [what you are watching, like the deployment or CI] finished and tell me what happened
Tip: The loop only runs while the session is open.
Step 5
Give it less
His advice: the minimal possible system prompt, the minimal possible tools, and let the model figure it out. Every tool and file that doesn't matter crowds the task.
Type /context to see what is filling the window, then drop what you don't need.
PromptClean the contextLook at /context and tell me what takes the most space and has nothing to do with [the task], and suggest what I should remove.
Step 6
Be open to being wrong, and /rewind
Boris says he is wrong so often now, and a lot of what he learned in engineering he has to throw out. So try new ways.
When things go wrong, /rewind rolls the code and the conversation back to a checkpoint, and you try again.
Tip: Files changed by Bash commands don't come back with /rewind, keep git with you.
Ready prompts to copy
Read CLAUDE.md and tell me the rules that matter for today's task: [the task]. Then start on it directly.
The takeaway
Let it work, let it learn from its mistakes, let it follow up for you. You focus on the task itself.
Most value
Start building with me now
We sit together, break down your problem, and I build the AI solution with you step by step
- A private session with you
- Real building during the session
- Results you take with you when we finish
100% free
A free consultation about your problem
The full path
Access to every guide (49), starting from 29 AED




