Day 3 – Skills Creation and Use Cases

1. CLAUDE.md – Audit & Upgrade

First thing we did – sent the agent a public template repository and asked it to compare with the current CLAUDE.md. The agent found gaps, proposed a plan, and upgraded the file from 21 lines to 66+.

Two levels: global (~/.claude/CLAUDE.md – for all agents on the machine) and project (.claude/CLAUDE.md – for a specific agent). Global = constitution. Project = job description.

Checklist

• Send the agent the template repo link it reads and compares on its own
• Review what the agent found missing sections, incomplete blocks
• Approve the upgrade the agent expands CLAUDE.md to 60–80 lines per the template
• Check AGENTS.md and TOOLS.md the agent creates stubs if missing
• Confirm exactly 4 @includes: USER.md, rules.md, decisions.md, handoff.md

Study this repository: https://github.com/qwwiwi/edgelab-claude-md

This is a reference guide for configuring CLAUDE.md. Compare it with my current CLAUDE.md:
1. Which sections are missing?
2. Which sections are incomplete?
3. What's redundant?

Then upgrade my CLAUDE.md to the reference level (60–80 lines). Show the result before applying.

Also check:
– Do AGENTS.md and TOOLS.md exist? If not – create stubs.
– How many @includes in CLAUDE.md? There should be exactly 4: USER.md, rules.md, decisions.md, handoff.md.

Sources

• edgelab-claude-md – full guide: global + project CLAUDE.md, templates, examples
• public-architecture-claude-code – reference architecture for memory, hooks, scripts
• docs.claude.com/claude-code/memory – official Anthropic documentation

2. USER.md – Operator Profile

Next, we filled in USER.md – the file where the agent learns who you are. Without it, the agent doesn't know your niche, channels, mission, and builds decisions "into the void." We recorded a voice message, sent social media links – the agent generated the file itself.

Checklist

• Record a voice message who you are, what you do, your mission
• Send links Instagram, Telegram channel, website
• Agent generates USER.md from voice + links
• Review and approve agent saves the file

Fill in USER.md – a profile about me. Here's my data:

Links:
– Telegram: [channel link]
– Instagram: [profile link]
– Website: [URL]

Who I am: [describe yourself in 2–3 sentences or send a voice message]

Generate USER.md using this template:
https://github.com/qwwiwi/edgelab-claude-md/blob/main/templates/user.md

Include: name, timezone, channels, mission, communication style, what I need from the agent. Show the result before saving.

Sources

• templates/user.md – template with placeholders

3. Rules.md – Agent Boundaries

Then we set up rules.md – operational boundaries. What the agent can do on its own, and what requires your permission. We started with 4–6 rules. The rest gets added organically – through work the agent learns and suggests new rules.

Checklist

• Green zone code, git, tests, small fixes (agent does autonomously)
• Red zone architecture, deploys, data deletion, force push (ask you first)
• Security never output keys, never commit secrets
• Git commits in your language, never push to main
• Escalation 1st attempt solo, 2nd with another model, 3rd – stop and report
• Output format "you communicate via Telegram bot, format for Telegram"

Fill in rules.md using this template:
https://github.com/qwwiwi/edgelab-claude-md/blob/main/templates/rules.md

Required sections:
1. Green zone (do autonomously): code, git commit/push to branch, tests, research, small fixes
2. Red zone (ask me): architecture, production deploy, data deletion, force push, spending > $50
3. Security: never output keys to stdout, never commit .env and secrets
4. Git: commits in [your language], branches feature/fix/refactor, never push to main
5. Escalation: 3 attempts, then stop and report
6. Output: you communicate with me via Telegram bot, format all messages for Telegram

Show the result.

4. OpenViking – Semantic Memory

Next step – the fourth memory layer (L4). OpenViking is an open-source semantic database. The agent searches by meaning, not keywords. Stores context across months and years. Without it, the agent only remembers the last 24 hours.

Two embedding options

Checklist

• Clone OpenViking to your VPS
• Create an OpenAI API key platform.openai.com, API Keys
• Set a spending limit Billing, Set hard limit ($5–10/month)
• Configure embeddings model text-embedding-3-small
• Test the search send a test query

Install OpenViking on the server for semantic memory (Layer 4).

Steps:
1. Clone the OpenViking repository
2. Configure embeddings via OpenAI API (model text-embedding-3-small)
3. Take the OpenAI API key from the OPENAI_API_KEY environment variable
4. Start the server
5. Send a test query: find "what did we discuss yesterday"
6. Show the result

If there's no OpenAI API key – explain how to create one at platform.openai.com and what limit to set.

Cron scripts for memory rotation

Without rotation, the HOT file bloats to 80KB+ per day – the agent starts ignoring instructions. 4 cron scripts solve this automatically.

Set up automatic memory rotation. Get the scripts from:
https://github.com/qwwiwi/public-architecture-claude-code/tree/main/scripts

Configure 4 cron jobs:
1. 04:30 – rotate-warm.sh (WARM older than 14 days → COLD)
2. 05:00 – trim-hot.sh (HOT older than 24h → compress via Sonnet → WARM)
3. 06:00 – compress-warm.sh (WARM > 10KB → re-compress)
4. 21:00 – memory-rotate.sh (COLD > 5KB → archive)

Verify the scripts work: run each one manually and show the result.

Sources

• public-architecture-claude-code/scripts – memory rotation scripts

5. Superpowers – Coding Skills

The installer already set up the Superpowers plugin – 15 skills that transform Claude Code from a chatbot into an engineer with processes. They activate automatically at the right moment, but you can call any of them manually.

Key Superpowers Skills

Try it right now

Send Jarvis a task and watch how Superpowers activate automatically:

Create a bash script health-check.sh that checks VPS health:
– Disk (df -h, warning if > 80%)
– RAM (free -h)
– CPU load (uptime)
– Service status: claude-gateway, claude-richard
– OAuth token: check that .credentials.json exists and is not empty

Output format: table [Check | Status | Details].
Save to ~/scripts/health-check.sh, make executable.
Use Superpowers: first plan, then TDD, then review.

What will happen: Jarvis activates writing-plans (plan), then test-driven-development (test before code), then verification-before-completion (verification). All automatic – you just watch.

Full list of skills

View all installed Superpowers:

ls ~/.claude/plugins/superpowers/skills/ | sed 's/\.md$//'

Documentation: github.com/pcvelz/superpowers

6. Reels Analysis & Skill Creation

We built a complete pipeline in real time: download Reels via HikerAPI, find the top 5 by views, transcribe via Groq, extract scripts and editor briefs. Then packaged everything into a skill via skill-creator.

Key principle: 50% scripts, 50% LLM. Scripts are like a crane that lifts cement to the 10th floor with one button. LLM is carrying it by hand every time.

Checklist

• Send the agent an account to analyze username or link
• Agent downloads the last 30 Reels via HikerAPI
• Finds top 5 by views and transcribes via Groq
• Generates an HTML report with scripts and editor briefs
• Package into a skill via skill-creator (prompt below)

Analyze the Instagram account @[username]. Work step by step:

1. Download the last 30 Reels via HikerAPI
2. Find the top 5 by views and comments
3. Download these 5 videos and transcribe via Groq
4. For each, extract: hook (first 3 seconds), script structure, CTA
5. Generate an editor brief: timecodes, teleprompter text, visual recommendations

Format the result as an HTML report and send it to me.

Use skill-creator. Package what we just did into a skill.

Requirements:
– Name: [slug in kebab-case]
– 50% of operations should be scripts (bash/python), not pure LLM
– Triggers: which phrases should activate it
– Input: what I provide (URL, username, topic)
– Output: what I get (HTML report, headline list, script)
– Tests: 2–3 examples for verification

Save to ~/.claude/skills/[slug]/SKILL.md

Tools

• HikerAPI – paid API for Instagram data ($0.08/request)
• Groq API – audio/video transcription
• skills.sh – skill-creator (official Anthropic skill, 164K+ downloads)
• fal.ai – image generation ($20 deposit, lasts 2+ months)

7. YouTube Producer Skill

The second skill we created live on stream. Take a channel, analyze the last 20–30 videos, extract the headline formula from top-performing videos, and generate new headlines + a script plan for your topic.

Analyze the YouTube channel [URL or name]. Work step by step:

1. Via YouTube API, get the last 20–30 videos (titles, views, date)
2. Find the top 10 by views in the last 2 months
3. Extract the headline formula: patterns, trigger words, structure
4. Download subtitles for the top 5 via Transcript API, extract script structure
5. Generate 10 new headlines for my topic: [topic description]

For each headline, suggest a brief video plan (3–5 points).

8. Skills Marketplace & Security

There are hundreds of skills on skills.sh. Each has a security audit: green means safe, red means don't install. Scammers leave skills that steal tokens. Always check the audit before installing.

Maximum 15–20 skills per agent. Once a week – cleanup: check logs, remove unused ones.

Checklist

• Find relevant skills via skill-finder or skills.sh
• Check the security audit before installing
• Install the agent installs on command
• Don't install more than 20 skills clean up weekly

Find skills for [my task] on skills.sh. For each, show:
1. Name and description
2. Security audit (risk level)
3. Download count

Install those that passed the audit (low risk). After installation, show my full skill list.

Go through the logs from the past week. Which skills did I never use? Show the list and suggest which ones to remove. Limit is 15–20 skills.

Sources

• skills.sh – skills marketplace for Claude Code

9. Product Development with Agent

Full cycle: from idea to working product. The senior-brainstorm skill generates the stack and architecture designed for 10,000 users. Then Superpowers handle code, tests, review. All through the agent.

Step 1. Install the senior-brainstorm skill

git clone https://github.com/qwwiwi/senior-brainstorm-skill.git ~/.claude/skills/senior-brainstorm

Step 2. Pipeline

1. Brainstorm (senior-brainstorm) – stack, architecture, trade-offs for 10K users
2. Specification – detailed requirements
3. Plan – Superpowers: writing-plans
4. Code – Superpowers: test-driven-development + executing-plans
5. Review – Superpowers: requesting-code-review + verification-before-completion
6. Deploy – staging, then production with your OK

Step 3. Launch

I want to create [product description]. Design for 10,000 users.

Use the senior-brainstorm skill: https://github.com/qwwiwi/senior-brainstorm-skill
If the skill is not installed – install it: git clone https://github.com/qwwiwi/senior-brainstorm-skill.git ~/.claude/skills/senior-brainstorm

Show me:
1. Stack (frontend, backend, DB, hosting, payments)
2. Architecture (2–3 options with trade-offs)
3. Implementation plan by phases

Consider: hosting on your preferred provider, data privacy laws in your jurisdiction, payment integration with your local payment processor.

After brainstorm – write a plan via Superpowers (writing-plans), then implement the first phase via test-driven-development.

What will happen: Jarvis activates senior-brainstorm (stack + architecture), then writing-plans (plan), then test-driven-development (code with tests). Superpowers activate automatically.

n8n integration

The agent can find and connect skills for n8n – no-code automations. n8n has an official MCP server, the agent manages workflows programmatically.

Find skills for n8n. What can they do? Can I use them to create automations?

Also check: does n8n have an official MCP server? If yes – connect it so you can build workflows yourself.