Small projects where an AI model or an agent did real work, with the pattern each one teaches. Concepts are in AI Agents and Workflows; tooling is in Claude Code Tooling and MCP.

Projects at a glance

Scraper with LLM Extraction

flowchart LR
    S[Scheduler: cron or Worker trigger] --> SC[scrape.py: fetch the public page]
    SC --> L[Claude: turn free text into JSON fields]
    L --> J[(data/results.json)]
    J --> R[render.py: build a Bootstrap page from the JSON]
    R --> H[Cloudflare Workers static assets]
    HC[healthcheck.py] -.-> J

What it taught: - Use the model only where rules fail. Fetching and layout are normal code; the model handles the irregular part (free-text terms → numbers) - Ask for structured output with a fixed schema; store null for anything missing instead of guessing, and show "n/a" - Keep the pipeline in stages (scrape, extract, save, render) so each can be re-run and tested alone - Store times in UTC, display local time (SAST) - Validate model output before saving (types, ranges) and run a health check for stale data - Static hosting on Cloudflare Workers is cheap and simple; a custom domain points at the Worker - Respect the source site's terms, rate limits and robots rules, and keep the source URL with each record

$ python scrape.py && python render.py
$ npx wrangler deploy

Hosting background in 3. Azure Networking and 5. AWS Networking (CDN, DNS ideas apply equally).

Source: sharifndlovu123/casino-bonus-scrapper (the Casino Welcome Bonuses scraper).

Website-Builder Pipeline (skills in sequence)

A test harness for iterating on design skills: put screenshots of sites you like into samples/<style>/, add a business document (PDF, DOCX or Markdown), then run the pipeline.

flowchart LR
    D[Business document] --> DD[doc-to-details: structure and content brief]
    SS[Reference screenshots] --> WD[website-designer: design analysis]
    DD --> CS[community-starter: build the static site]
    WD --> CS
    CS --> PA[production-audit and content-integrity-audit]
    PA --> OUT[output/ preview locally]

Lessons: - Chain small skills with clear files between them (details.md, design-render.md) so each stage is inspectable - Audit against the source: a content-integrity check catches invented stats, wrong contact details and dead links - Audit for production: semantic HTML, image and font handling, caching and security headers, host config (_headers, _redirects) - Keep a context/, errors/ and docs/ folder so failures become learnings

Programmatic Video with Remotion

Remotion renders React components as video frames, so a video is code: scenes, timing, data and styling in TypeScript.

$ npm i
$ npm run dev          # preview in the Remotion Studio
$ npx remotion render   # render to MP4

• A Composition defines size, fps and duration; scenes are components
• Animation is driven by the current frame (useCurrentFrame, interpolate, spring)
• Great for data-driven videos, repeatable templates, and generating many variants
• An agent can edit scenes and constants quickly because everything is plain code

n8n Automation with Django

n8n is a workflow automation tool with a visual editor and an API. A Django app can create workflows programmatically and trigger them, so business processes (case management, approvals, notifications) run as n8n flows while Django owns the data.

$ npm install -g n8n
$ n8n start              # UI on http://localhost:5678, then create an API key

Pattern: Django model → WorkflowManager calls the n8n API → n8n runs steps and calls back with webhooks. Compose runs Django, PostgreSQL and n8n together (2. Docker Compose). Comparison with BPMN engines is in BPMN and Workflow Engines.

Putting an Agent in the Loop

Where AI helped most in these projects: 1. Turning messy input into structure (scraping, documents) 2. Producing first drafts of specs, sites and code from a clear brief 3. Reviewing work with a fresh, independent reader 4. Repetitive, well-specified edits across many files

Where a human stays in charge: approving designs, merging code, deploying, anything with money or customer data.

Writing a Book with Claude Code

sharifndlovu123/book-of-tartaros uses Claude Code as a long-running collaborator for prose and continuity across many sessions. The repo itself is the collaboration, and the same habits that keep code projects on track work for long writing:

• A canonical "bible" (worldbuilding, characters, structure) is the source of truth; everything else is drafting and reference
• A questions tracker records open character and plot questions until they are closed
• Dialogue style cards give each voice a consistent sound
• A CLAUDE.md project brief tells the agent the rules at the start of every session
• Chapters live as separate Markdown files, with a consolidated full-book file built from them
• Drafts and scratch work are git-ignored, so only the settled text is tracked

The pattern generalises: write the rules and facts down once in a file the agent reads every time, instead of relying on chat history (AI Agents and Workflows).