DEV Community: Nex AI

I Built 15 OpenClaw Skills for Belgian SMEs: Technical Decisions and Lessons Learned

Nex AI — Sun, 05 Apr 2026 21:25:48 +0000

Last month, I analyzed the ClawHub registry and realized something striking: 13,700+ skills available, but business finance categories were nearly empty. No skills for Belgian e-invoicing, no VAT tools, no SME-specific workflows. As founder of Nex AI, a Belgian digital transformation agency, I saw an opportunity. I built 15 skills to fill that gap in one month. Here's what I learned.

The Gap Analysis: Finding the Empty Shelves

ClawHub's registry is massive, but the Pareto principle applies hard. I indexed the 13,700 skills and grouped them by category. Automation, AI, and developer tools dominated. But look at business processes? Finance? Compliance? Crickets. Worse, anything Belgium-specific was nonexistent.

For an SME operator in Belgium or Flanders, there were no skills to handle:

Mandatory e-invoice format changes (happened Jan 2026)
VAT calculation for digital services
Freelancer invoice templates
Agency retainer tracking

This gap made sense. ClawHub wasn't saturated in vertical niches. Most builders were chasing broad audiences. But that meant zero competition for Belgium-specific, finance-adjacent tools that my network actually needed.

I decided to ship 15 skills in one month. All Python. All simple. All targeting freelancers, SMEs, and agency operators.

The Belgian E-Invoicing Mandate: Technical First Priority

On January 1, 2026, Belgium enforced the Peppol BIS 3.0 standard for e-invoicing. UBL 2.1 XML format. Overnight, any freelancer or small agency invoicing a Belgian company needed compliant invoices or face penalties.

No existing skill handled this. I built nex-einvoice first.

The design decision was straightforward: Python only, stdlib only, zero external dependencies (except optional tesseract/whisper for OCR, optional ffmpeg for media). Why? Because ClawHub skills are distributed as single Python files. No pip install. Users drop the file, run it, done. Dependencies break that promise.

For e-invoicing compliance, I needed to:

Parse invoice data (name, amount, dates, VAT)
Generate valid UBL 2.1 XML
Validate against Peppol rules
Output a file ready to send to a Belgian invoicing hub

Python's xml.etree.ElementTree handles XML generation without dependencies. The stdlib datetime and decimal modules ensure precision. The entire skill is ~800 lines.

Technical Architecture: Consistency Across 15 Skills

I needed consistency. All 15 skills follow the same patterns:

SQLite with FTS5 everywhere. Every skill that needs persistent storage uses SQLite. Why? It's bundled with Python. No server. No migrations. FTS5 (full-text search) lets me index invoice descriptions, client names, VAT codes, whatever. Query time is milliseconds. File is portable, versioned in Git.

Argparse CLI across all 15. Every skill uses argparse for CLI parsing. Users run:

python nex-einvoice.py --invoice-file invoice.csv --output peppol.xml --validate

Consistent interface. Easy to chain into workflows. Every skill documents its arguments the same way.

Shared patterns, DRY principle. All 15 skills use:

_check_db(): Initialize SQLite with schema if needed, idempotent
_db_conn(): Context manager for safe connections
FOOTER constant: Consistent markdown output footer with GitHub link
lib/ modules: Shared validation logic, formatters, e-invoice builders

Example SKILL.md frontmatter (ClawHub format):

name: nex-einvoice
description: "Generate Belgian-compliant e-invoices in the Peppol BIS 3.0"
UBL format from natural language input in Dutch or English...
(150-300 words, this is your entire SEO on ClawHub)
version: 1.0.0
metadata:
clawdbot:
emoji: "\U0001F9FE"
requires:
bins:
- python3
env: []
homepage: https://nex-ai.be
files:
- "nex-einvoice.py"
- "lib/*"

- "setup.sh"

Note the MIT-0 license (separate LICENSE.txt file). ClawHub requires MIT-0 or compatible. Non-negotiable.

The 15 Skills: Grouped by Tier

Tier 1: Core Business Tools (5 skills)

nex-einvoice: Peppol BIS 3.0 / UBL 2.1 e-invoice generator from natural language
nex-expenses: Receipt OCR and expense categorization with 15 Belgian tax deduction buckets
nex-crm: Chat-native prospect CRM with pipeline management
nex-vault: Contract and document vault with auto-renewal detection and expiry alerts
nex-healthcheck: Multi-service monitoring dashboard with 9 check types

Tier 2: Agency Operations (5 skills)

nex-deliverables: Client deliverable tracker for agencies managing multiple projects
nex-domains: DNS and domain portfolio manager with Cloudflare API sync
nex-reports: Scheduled report generator aggregating data from 10 modules
nex-keyring: API key rotation tracker for secret hygiene and audit compliance
nex-changelog: Release notes generator with git parsing and client email formatting

Tier 3: Specialized Tools (5 skills)

nex-voice: Voice note transcription with Whisper integration and action item extraction
nex-pricewatch: Competitive price monitor with CSS/XPath/regex scraping
nex-onboarding: Client onboarding checklist with 20-step agency workflow template
nex-gdpr: GDPR/AVG data request handler for Articles 15-21 compliance
nex-skillmon: Skill health and cost monitor for OpenClaw power users

What I Learned About ClawHub Publishing

ClawHub uses semantic vector search on the description field only. Your title is nice, but your description IS your SEO.

I made mistakes on early submissions:

First version of nex-einvoice: 45 words, stuffed keywords. Poor ranking.
Revised: 150 words, natural language, explained the Belgian mandate and Peppol standard. Top results for "belgian invoice" and "peppol xml".

Sweet spot: 150-300 words. Write for humans, not search bots. Explain the problem, the solution, who benefits. ClawHub's algorithm finds you.

One more thing: early 2026, ClawHub had a malware crisis. Someone published 200+ fake skills (crypto scams, data theft). Trust matters. I published all 15 skills open-source on GitHub: github.com/Nex-AI-Guy. Every skill is auditable. Zero external dependencies. I link to GitHub in every SKILL.md. Transparency is a feature, not a burden.

The Business Model: Free Skills, Paid Services

This is the lead funnel. The 15 skills are free, MIT-0 licensed, open-source. They solve real problems for Belgian freelancers and SMEs. Users download, use, appreciate. Some will ask: "Can you customize this? Can you integrate this into our workflow? Can you train our team?"

That's where Nex AI enters. Retainer services. Deeper integrations. Custom skills. The free skills are the front door.

What's Next

All 15 skills are live on ClawHub now. I'm tracking which descriptions perform best, which search terms drive installs, and which skills generate inbound interest. Early signs are promising for the Belgian e-invoicing niche since there's literally zero competition.

If you're building skills for a specific geography or vertical, you already know the gap. You see the empty shelves. The market is there. Start with the simplest problem (the e-invoicing mandate was my forcing function), build backwards, and ship.

Explore all 15 skills on ClawHub (@NexaiGuy), or check out the source code at github.com/Nex-AI-Guy. I'm curious what gaps you see in your market.

Kevin Blancaflor is founder of Nex AI, a Belgian digital transformation agency based in Gent. Find all skills at clawhub.ai/@NexaiGuy.

I Built an AI Agent Skill That Remembers Everything You Do on Your Computer

Nex AI — Sat, 04 Apr 2026 07:53:15 +0000

AI agents are powerful. But they have a blind spot: they start every conversation from zero. They don't know what you were researching yesterday, which YouTube tutorial explained that technique you need, or how you solved a similar problem last month.

I wanted to fix that. So I built Nex Life Logger, a background activity tracker that gives your AI agent persistent memory of what you actually do on your computer. It tracks your browsing history, active windows, and YouTube videos, stores everything locally, and lets you query it all through natural language.

67,000+ activities tracked so far. Zero data sent to the cloud.

The Problem

I run a digital transformation agency in Belgium (Nex AI) and spend my days jumping between code editors, documentation, design tools, YouTube tutorials, and dozens of browser tabs. By Friday, I've forgotten half of what I researched on Monday. I needed a way to ask "What was I working on when I was debugging that Docker issue?" and get an actual answer.

Existing tools like Rewind and Windows Recall either send your data to the cloud or lock you into one platform. For someone who cares about where their data goes, that was a dealbreaker.

How It Works

Nex Life Logger has a simple architecture. A background collector runs every 30 seconds and captures three things:

Browser history from Chrome, Edge, Brave, and Firefox. It copies the locked SQLite history files to temp, reads them, then securely deletes the copies (overwritten with random bytes before unlinking). No browser extension needed.

Active window focus. It checks which application and window title are in the foreground. This gives context about what tool you were using, not just what you were browsing.

YouTube transcripts. When you watch a productive video, it automatically fetches the full transcript and stores it. This is where the real value is. When the AI generates summaries, it can tell you exactly what was discussed in each video, including the key concepts, tools, and techniques mentioned.

Everything lands in a local SQLite database at ~/.life-logger/. WAL mode for crash safety, owner-only file permissions, no external connections.

The Filtering Layer

Tracking everything would be a privacy nightmare. So there are multiple filtering layers:

Chat and messaging apps are excluded automatically. WhatsApp, Discord, Slack, Telegram, Signal, Teams, and anything with /messages, /chat, or /dm in the URL. Sensitive windows too: password managers, banking apps.

There's also a productivity filter. The system only tracks content related to AI, programming, design, building, and learning. Politics, news, entertainment, sports, and celebrity gossip are automatically filtered out. You can configure all of this through 12 customizable filter categories.

AI-Powered Summaries

Raw activity data is useful but noisy. The real power comes from hierarchical AI summaries.

Every night at 11 PM, the system generates a daily summary from that day's raw activities and transcripts. On Sundays, it generates a weekly summary from the daily ones. On the 1st of each month, a monthly summary from the weeklies. On January 1st, a yearly summary from the monthlies.

Each level builds on the one below it, creating a compressed but rich record of everything you've done. After a few weeks, you have a layered memory system that an AI can search through instantly.

The summarization works with any OpenAI-compatible API. I'm running it with Alibaba's Qwen (free tier), but you can use OpenAI, Groq, Ollama for fully local, or any compatible endpoint.

Querying Your History

There are two ways to query your data.

The desktop app has a built-in AI Search tab. Type a question like "What was I researching about Docker last week?" and it searches across activities, summaries, keywords, and transcripts, then generates an answer with specific dates, links, and context.

The OpenClaw/ClawHub skill lets you query through any connected agent, whether that's Telegram, Discord, or WhatsApp. I published the skill to ClawHub (the OpenClaw skill registry) so anyone running OpenClaw can install it with one command:

npx clawhub install nex-life-logger

Then you message your agent naturally. "What was I working on yesterday?" and it calls the CLI, searches your local database, and answers in the conversation.

The CLI

The ClawHub skill includes a full CLI for querying your data directly:

nex-life-logger search "docker containers"
nex-life-logger summary daily
nex-life-logger activities --last 2h --kind youtube
nex-life-logger keywords --category tool --top 15
nex-life-logger stats

There are commands for searching, browsing activities, viewing summaries, extracting keywords, managing the background collector, configuring AI providers, and exporting your data in JSON, CSV, or HTML.

What I Learned Building It

SQLite is perfect for this use case. Fast writes, zero configuration, single file backups. With WAL mode, the collector and the viewer can read and write simultaneously without conflicts.

Content filtering is harder than expected. The productivity filter went through multiple iterations. Pure keyword matching catches obvious cases but misses nuance. I added an optional AI-based classifier for ambiguous content, but kept it off by default to avoid unnecessary API costs.

YouTube transcripts are gold. The daily summaries became 10x more useful once they included what was discussed in videos. Instead of "watched 3 YouTube videos about Docker," you get "watched a tutorial on multi-stage Docker builds covering layer caching, build arguments, and image optimization for production."

Privacy-first isn't just a feature, it's a design constraint. Every decision was filtered through "does this keep data local?" No cloud sync, no telemetry, no analytics. The API key is stored in Windows Credential Manager, not a config file. Browser history files are securely deleted after reading. It adds complexity, but it's the right tradeoff.

What's Next

The current version tracks browsers and windows. I'm considering adding:

Clipboard history (what you copy and paste throughout the day)
Terminal commands (your shell history with context)
File system changes (what files you created, modified, deleted)
Calendar integration (correlate activities with meetings)

Each one adds another dimension to the agent's memory of your workday.

Try It

Desktop app (Windows, macOS, Linux):
GitHub: NexaiGuy/nex-life-logger

OpenClaw/ClawHub skill (one-command install):

npx clawhub install nex-life-logger

ClawHub: nex-life-logger

Built by Nex AI, a digital transformation agency in Belgium. If you work with AI agents and want this kind of integration for your business, that's what we do.

What data sources would you want your agent to remember? I'd love to hear what would be most useful.