ابنِ مساعدك الشخصي على Claude Code بأربع قطع: المهارات والذاكرة والصوت والموزّعBuild Your Own Assistant on Claude Code in Four Parts: Skills, Memory, Voice and a Router

Key terms

skill

A SKILL.md file holding a task written down once. Run it with a command like /morning-brief, or Claude runs it when your request matches its description.

vault

An Obsidian notes folder on your computer. Plain Markdown files that Claude can read and write.

CLAUDE.md

A file in the folder that tells Claude where everything is: where to look and where to save its work.

router

A small step before the model that decides whether a request needs a big model, a fast cheap one, or no model at all.

The steps

1. 1

   Step 1

   Find the skills in your own work

   Start from what you do every day, not from a new tool. Open Claude Code in your work folder and ask it what you repeat.

   The rule is simple: anything you did by hand twice becomes a skill.

   PromptFind your skills

   Copy

   From my work in this folder over the last 30 days, which tasks do I repeat? Suggest 5 clear skills, each with a short name, what it does, and when I run it.

   Tip: Start with three skills. One skill used daily beats ten nobody touches.

2. 2

   Step 2

   Write your first skill

   Each skill is a folder with a SKILL.md inside. Personal skills live in ~/.claude/skills/<name>/SKILL.md and work in every project.

   At the top, write a clear description, because Claude reads it to decide when to use the skill. Below it, write the instructions.

   code.claude.com

   

   PromptMake the skill

   Copy

   Create a skill called [morning-brief] in ~/.claude/skills. Description: [a morning brief from my notes and calendar]. Instructions: [read raw/ and today's note, write 5 points]. Save the result in outputs/.

   Tip: Claude Code picks up skill edits in the same session. No restart needed.

3. 3

   Step 3

   Set up the memory in Obsidian

   Obsidian is free and keeps your notes as Markdown files on your computer. Make a vault with three folders: raw/ for everything you capture, wiki/ for organised notes, outputs/ for what Claude produces.

   Then write a CLAUDE.md at the root that explains this layout. That is the map Claude follows.

   PromptWrite the map

   Copy

   Write a CLAUDE.md for my folder. It has raw/ (everything I capture), wiki/ (organised notes) and outputs/ (what you produce). Tell me where you look first, where you save any report, and never edit anything in raw/.

4. 4

   Step 4

   Add the voice: it hears you and answers

   faster-whisper turns your speech into text on your computer, and it is open source. Kokoro turns the reply into speech and runs locally once its model files are downloaded.

   Note: Whisper understands Arabic, but Kokoro has no Arabic voice yet. For a spoken Arabic reply you need another speech tool.

   PromptThe voice script

   Copy

   Write a Python script that records from the mic until I stop, transcribes with faster-whisper on CPU (int8), and sends the text to claude -p. List the install steps before the code.

5. 5

   Step 5

   Put a router in front of the model

   Not every request needs Opus. "Open my morning brief" needs no model at all, just a keyword match. "What's the news today?" is fine on Haiku. "Make me a report" is where full Claude Code runs.

   Start with simple rules: fixed keywords first, then a light model, and the big one only for real work.

   PromptBuild the router

   Copy

   Write a route(text) function with three tiers: 1) if the text has a word from [the list], open the file directly with no model. 2) A short question goes to Haiku. 3) Any request with make, write or analyse goes to claude -p. Return the tier name with every answer.

   Tip: Print the tier with every answer in week one so you see where the router gets it wrong.

6. 6

   Step 6

   Run a skill without a chat

   claude -p runs Claude Code without the interface: you send a request, it does it and prints the result. That is what lets a button or a voice command run a whole skill.

   When you run it from a fixed script, add --allowedTools with only the tools it needs.

   code.claude.com

   

   PromptThe run command

   Copy

   claude -p "/morning-brief" --allowedTools "Read,Write"

7. 7

   Step 7

   Ask for the UI, and keep sensitive actions behind your approval

   The last part: a page with a button per skill and a place to read outputs/. Ask Claude Code to build it.

   Start with skills that only read and summarise. Any skill that sends email, pays or publishes should ask for your approval before it acts.

   PromptMake the UI

   Copy

   Build a simple local web page: one button per skill in ~/.claude/skills, each button runs claude -p "/<skill name>" and shows the result, with the latest files in outputs/ listed below. No login, localhost only.

Ready prompts to copy

Review my skills in ~/.claude/skills and the outputs/ folder from this week. Which skill never ran? Which one needs changes? Suggest one new skill from something I repeated.