Builders who want to use Claude for personal projects get the most out of it by starting with a specific annoyance, not a technology. Felix Rieseberg, an Anthropic engineer who built Claude Cowork, proves this with three side builds: a 3D floor plan walkthrough tool, an automated promise tracker, and a $20 hardware AI buddy that calls the Claude API over Wi-Fi. Copy the pattern, not the projects.
In the 2025 Stack Overflow Developer Survey, 76% of professional developers reported using AI coding assistants for personal side projects, up from 55% in 2023. That shift is real. Claude for personal projects now works at a friction level that once required a full team. The key move is treating Claude as an interpretation layer: messy real-world input goes in, clean actionable output comes out. Rieseberg's three builds all follow that shape. So can yours.
Who Is Felix Rieseberg and What Is Claude Cowork?
Felix Rieseberg is a senior engineer at Anthropic with a long track record in developer tooling. Before Anthropic, he shipped Electron apps and open-source projects used by millions of developers. His current project at work is Claude Cowork, a multiplayer AI workspace that lets several users share a live, AI-assisted session inside Claude. Think of it as a shared canvas where the model sits at the table from the start, not summoned later.
That background matters here. Rieseberg is not a developer who dabbles with Claude on weekends out of curiosity. He builds the surfaces other engineers use to ship with it. When someone at that level reaches for Claude on a personal side project, the choice tells you something honest about where the tool actually earns its keep.
In an interview with Lenny's Newsletter, Rieseberg walked through three personal builds in detail. Each one started with a real friction point he already felt in daily life. None started with "let me see what this AI can do." All started with a problem. That ordering turns out to matter a lot.
How Does He Turn Floor Plans into 3D House Walkthroughs?
The workflow is simpler than it sounds. Rieseberg uploads a 2D floor plan image, prompts Claude to read the room labels and estimate dimensions from the visual layout, and receives structured code output: Three.js or Blender Python that a renderer can run directly. As of May 2026, Anthropic's Claude API supports vision inputs natively across all model tiers, so no separate OCR preprocessing step is needed before the image hits the model.
The output is a navigable 3D model. It is not an architect-grade rendering. Walls stand, rooms connect, and you can walk the space in a browser. That is enough to spot layout problems before a contractor visit or a furniture order.
Practical limits apply. Accuracy drops on hand-sketched plans that lack labeled measurements. The model infers scale from context, which works well on standard printed floor plans and poorly on napkin sketches with no scale reference.
For a live screen demo of the 2D-to-3D transformation, the Lenny's Newsletter interview includes a walkthrough. The GenAI Club team is running a replication test with a sample floor plan; confirmed output details will update this section once the build is complete.
What Is Automated Promise Tracking and Why Is It Hard to Solve Otherwise?
Promises scatter. You tell someone "I'll send that by Friday" over Slack. You commit to a follow-up in an email thread. You mention a deliverable in a standup. None of those turn into calendar items automatically. Off-the-shelf task managers fail here because they require you to log a commitment at the exact moment you make it. That rarely happens under real work conditions.
Rieseberg's solution uses Claude to read communication threads and flag unresolved commitments without manual entry. Claude identifies language patterns that signal a promise: "I'll send," "let me get back to you," "I can handle that." It surfaces them as open items.
As of May 2026, Claude's Projects feature supports persistent memory across sessions, making this kind of tracking meaningfully more reliable than when the feature launched in 2024. A running commitment log inside a Project builds up over time without re-uploading context each session. If you want to set up a similar system, this guide on using Claude Projects as a knowledge base walks through the structure step by step.
How Does the $20 Hardware AI Buddy Actually Work?
The parts list is short. A Raspberry Pi Zero 2W runs roughly $15. A USB microphone costs about $5. A small speaker connects through the audio jack. As of May 2026, the full component set lands under $25 at major electronics retailers, down from roughly $40 for equivalent hardware in early 2024.
The device runs a lightweight Python loop. It records a spoken question, sends the audio to a transcription service, forwards the transcript to the Claude API, and plays the response through the speaker. No screen. No app install. You ask; Claude answers.
Commercial AI hardware devices sit above $100 at retail. This build does the same core job for a fraction of that, and you control every layer of the stack.
For technical readers: The GenAI Club team is running a replication build with a Raspberry Pi Zero 2W to document actual component costs, setup time in minutes, and failure points. Results will update this section once the test is complete. Use the cost estimates above as a planning baseline. To keep API spend low on a low-power device, this breakdown on reducing AI API costs is worth reading first.
What Pattern Connects All Three of These Projects?
Look at the three builds side by side and the same shape appears. Each starts with a friction point Rieseberg already felt in daily life. Not a technology searching for a problem. A problem that needed a minimal solution.
Claude sits in the middle as an interpretation layer. A 2D floor plan image goes in; structured 3D code comes out. A scatter of Slack messages goes in; a list of open commitments comes out. A spoken question goes in; a spoken answer comes out. In each case, Claude converts messy real-world input into something immediately usable.
The low iteration cost changes behavior. A failing prototype costs minutes, not days. You test whether an idea works before committing to building it. That low-stakes testing is what separates builders who ship personal AI tools from those who stay in planning mode.
The same input-interpret-output pattern scales beyond personal projects. The Model Context Protocol: How MCP Connects AI to Your Tools guide covers how MCP formalizes this architecture when Claude needs to connect to external tools and data sources at production scale.
How Do You Apply This Approach to Your Own Projects?
Start with a friction log. Spend one week writing down every manual task that feels repetitive or error-prone. Do not filter for "AI-solvable" ideas at this stage. Just log friction. After seven days, look for patterns where the work is mostly translating something messy into a clean action. Those are your candidates.
Test in Claude.ai before writing any code. Most ideas succeed or fail at the prompt stage. Upload a real sample of your messy input, write a prompt that asks for the output you want, and see what returns. This costs nothing and takes five minutes.
Know when to stop. If a prompt needs more than ten minutes of refinement for a simple task, the problem framing is wrong, not the tool. Restate the problem from scratch and try again. The model is rarely the bottleneck at this stage; vague inputs produce vague outputs every time.
For a longer look at Claude workflows that hold up over time, five automation workflows that survived six months in production gives an honest field report from a builder who tested each one under real conditions. And if you want to sharpen the prompts that drive any personal project, these prompt engineering techniques that actually work in 2026 gives a practical starting set.
The most durable Claude for personal projects work shares one trait: it starts with a real annoyance, not a technology. Rieseberg's floor plan tool, promise tracker, and hardware buddy each solve a specific friction point by using Claude as the interpretation layer between messy real-world input and clean, actionable output. Find your friction first. Test the prompt next. Build last.
If you want to see practical AI workflows built live, Gen AI Summit Asia is opening in Kuala Lumpur on August 8-9, 2026: two days of AI shortcuts across eight real business tracks. Find out more about Gen AI Summit Asia.
FAQ
Can Claude actually read a floor plan image and turn it into a 3D model?
Yes, with real limitations. Claude's vision input can interpret a 2D floor plan image, extract room labels and approximate dimensions, and generate code (typically Three.js or Blender Python scripts) that produces a navigable 3D representation. The quality depends heavily on how clearly the floor plan is labeled. Scanned blueprints with explicit room dimensions and a printed scale work well. Hand-sketched plans without measurements produce rough, sometimes inaccurate approximations. Felix Rieseberg demonstrated this workflow in a Lenny's Newsletter interview, using it to evaluate a house layout before a purchase decision. The output is functional for spatial reasoning, not a substitute for architect-grade rendering software.
What is Claude Cowork and who built it?
Claude Cowork is a multiplayer AI workspace that runs inside Claude, letting multiple users collaborate within the same AI-assisted session and share context in real time. It was built by Felix Rieseberg, a software engineer at Anthropic. The project demonstrates how Claude can serve as a shared context layer for team workflows rather than only a solo assistant. Rieseberg built it as both a practical internal tool and a proof of concept for real-time AI collaboration. As of May 2026, details on public availability can be found through Anthropic's official communications and Rieseberg's public posts.
How does AI promise tracking work in practice?
Promise tracking uses Claude to monitor text-based communication streams such as emails, Slack messages, or meeting notes, and identify statements that represent a commitment. Phrases like 'I will send that by Friday' or 'we will follow up on this' get flagged as open promises. Claude can resurface these when a deadline approaches or when a thread closes without a resolution recorded. Rieseberg's version connects to his personal communication channels rather than a separate task manager, so commitments are captured in context rather than requiring manual entry. The main limitation is that it works only on text; verbal-only commitments still require a transcript before Claude can act on them.
How do I build a $20 AI hardware buddy using the Claude API?
The core build uses a Raspberry Pi Zero 2W (roughly $15), a USB microphone (roughly $5), and a small speaker. The device runs a Python script that records voice input, sends the transcript to the Claude API over Wi-Fi, and plays the response through the speaker. Total component cost lands between $20 and $25 depending on sourcing. You need an Anthropic API key, which bills per token. Setup steps: flash Raspberry Pi OS Lite, install Python and the Anthropic SDK, write a record-transcribe-respond loop, configure the script to run on boot. Someone comfortable with basic Linux terminal commands can have a working prototype in two to three hours.
What is the best way to find personal Claude project ideas?
Start with a friction log rather than a feature list. For one week, write down every manual task that feels repetitive, error-prone, or slow. Then ask whether the bottleneck is interpretation (turning messy input into structured output) or execution (acting once you know what to do). Claude excels at interpretation: reading documents, categorizing information, generating code, and converting formats. Rieseberg's three projects all follow this pattern. Each one addressed a specific annoyance in daily life. Starting from the technology ('I want to use Claude for something') reliably produces weaker projects than starting from the friction ('I hate doing this manually every week').
Is building personal tools with the Claude API expensive?
For most personal projects, the ongoing cost is low. Claude's API pricing as of May 2026 is token-based, and lightweight personal tools running a few dozen queries per day typically cost under five dollars per month. The $20 hardware buddy is a good reference point: hardware is a one-time cost, and casual daily interaction stays well under two dollars per month at typical usage levels. Costs scale if you process large documents frequently or run many concurrent sessions. For prototyping, Claude.ai's free tier is sufficient to test most personal project ideas before committing to any API spend.
What skills do I need to replicate these Claude projects?
The floor plan and promise-tracking projects require basic prompt engineering and the ability to test code Claude produces, but not the ability to write code from scratch. Claude generates the scripts; your job is to run them, spot what breaks, and iterate. The hardware buddy requires more hands-on comfort: setting up a Raspberry Pi, running terminal commands, and editing a Python file. Rieseberg is a professional software engineer, but all three projects are deliberately simple in architecture. A motivated beginner with basic command-line familiarity can replicate the hardware build over a weekend. The floor plan project needs nothing beyond a Claude.ai account.