Last week I left you with a nail-biter. I ran a sneaky wifi network near a weird marathon in 2018 and I captured close to 200 devices. I reproduced the experiment this fall- how’d it go in 2025? Terrible in some regards, but awesome in terms of prototyping acceleration. An experiment that took 2 months in 2018 took me 4 days in 2025.
The Bad
In the 2025 experiment, I caught a grand total of 18 devices.
Does this mean mobile phones are more secure? Was it the exact same experiment? No!
Low Participant Turnout: My Wifi Hotspot was active starting at 7 am. The marathon was scheduled to start at 8 am. We didn’t see a single runner till ~9:15 am. When runners did start arriving, the quantity of runners was limited compared to past years. The 2018 marathon spanned two days. The race was only one day this year. The participant cohort of runners was significantly smaller than in years past.
Bad SSID choices: This attack depends upon your ability to anticipate a wifi ssid that your targets have an affinity for. The wifi SSID i used in 2018 wasn’t going to work because it has been deprecated. I went with “Starbucks WiFi” initially, but this only caught 2 devices. The lack of “Starbucks WiFi” tuned devices is an interesting indicator of how times have changed. It used to be that mobile phone owners needed to attach to wifi to use email/browse the web with their phones. This was because cellular networks did not have unlimited data, and so you either ran out of data for the month or you were hit with a large cellphone bill if you used cellular for data. People used to go to coffee shops to “work” on their phones and laptops. Now you’re really there to socialize or caffinate. I also wonder if Starbucks’ popularity has declined. In the last 10 years, I’ve only drank Starbucks out of necessity.
So after a couple of hours of watching only 2 attaches, I yielded to temptation and changed the SSID to “xfinitywifi.” The xfinitywifi ssid is a controversial wifi network vended by Comcast, exclusive to comcast customers.
You can use wigle.net to see the most popular active SSIDs:
Changing to use xfinitywifi felt like desperation! Comcast does not have much presence in Snoqualmie valley. I reasoned that most of the runners were probably coming from cities where Comcast is dominant- e.g. Bellevue, Issaquah and Redmond. I managed to catch 16 more devices over the next 4 hours. The count was so small I didn’t bother to keep my logs. But here are some screenshots to give you a feel for what I experienced:
This experiment agitated me greatly. I know there are still problems related to wifi offloading- but I only caught 18 devices. I didn’t spend enough time researching SSIDs and the end result was low attaches.
Despite my grumpiness about the data, this experiment was a major success.
Did you notice the external Wifi Adapter above? How about the nice Paperwhite display presenting status of the device. My monitoring script was far more sophisticated than a tail of a hostapd logs. I didn’t have to write this code or fiddle with hostapd configurations or nftables rules. I didn’t have to find the right kernel headers and compile wifi drivers. I didn’t have to flex my terrible design skills. I knew the features I wanted and I gave my agents direction on how to deploy the features.
I was able to successfully produce an IoT prototype with complex hardware dependencies in 4 days.
The Good:
I implemented a working prototype of a custom wifi hotspot with a paperwhite display, an external wifi adapter & a Yagi wifi antenna in 4 days.
Methodology
Claude Code & Pre-prompting strategies
I leveraged Claude code for most of my work. I created a working directory invoked Claude with a 1,500 line pre-prompt for requirements analysis and planning. This pre-prompt produced ansible playbooks that take advantage of my Firmware Development caching containers. The pre-prompt addresses topics related to Requirement Exploration, Architecture Safety, Known Good Deployment Patterns, Domain Specific Knowledge and Documentation & Maintenance. I’ve been iterating on this prompt for about 6 weeks through applications on about 5 other projects. I constructed a separate pre-prompt of 166 lines that handles Deploying code, Code analysis, system access, frameworks for deploying code & systematic troubleshooting and refactoring the code to address discovered defects.
Development Loop
The normal lifecycle of developing a reliable working prototype seems to take about 3-4 build cycles.
My agent would serially perform the following operations during the build process:
- Initiate a build
- Discover defects during build process
- Troubleshoot them on the recipient system
- Make corrections to the original build playbooks
- Resume the build at the corrected defect
- Complete a working build.
If the build experienced errors, I waited to get a complete build and then started again on fresh recipient image. I kept seeing improvements until the build process ran reliably without errors.
Throttling
My biggest challenge was rate limiting:
My agents hit my 5 hour Anthropic token limit on the $20 plan in about 2 hours. During this 4 day period, I scheduled my day around throttling limits. I tried to make sure that some building happened while I slept. Two days before the marathon, I upgraded to the $200 plan. My iOS screen time report was 1 hour during that week.
I didn’t have to write any code to make this project work. That’s not to suggest that anybody could do this experiment. I was successful because I knew exactly what software libraries I wanted to see deployed and how I wanted them tuned. I regularly had to intervene when the agents proposed bad plans. But I’m now approaching a point where my single board computer development processes are automated. It felt like having a mildly competent apprentice.
Over the last few years, I’ve been able to build a range of Raspberry Pi Prototypes. All of them were a labor of large effort. My build process made prototyping faster, but it still took me several months to work out the details of various project:
- An rPi turned into a WiFi Hotspot that can be used for learning WiFi hacking. This enables kids to safely learn WiFi security concepts.
- An rPi that works as a wifi bridge for an existing Ethernet network. This enables me to expand my wired Ethernet network to my workshop inexpensively. Normal humans buy expensive WiFi access points. I build them from scratch!
- An rPi with a self hosted Ghost Blog & Planka Kanban board + cloudflared tunnels. This enables me to track my progress on personal goals and projects without paying money to Amazon or Trello.
Making reproducible builds was expensive and typically took 2-3 months. I’d steal spare time on evenings or weekends to work on projects. The greatest costs come from the testing & validation needed to create durable, reproducible firmware images. With a combination of tasteful pre-prompts, custom agents & an automated build process I can turn around reproducible firmware builds in less than a week.
1. Software & Hardware Testing Houses
You need repeatable, cost-effective environments to validate new software and hardware under real-world conditions, but setting up and tearing down test rigs is slow, inconsistent, and prone to configuration drift.
2. Managed Security Service Providers (MSSPs)
You need deployable, trusted network nodes inside customer environments for monitoring, detection, and incident response — but sourcing, configuring, and reproducing reliable hardware platforms across dozens of clients eats up valuable engineering time.
3. IoT Manufacturers
You want to prove out your next device concept quickly, with working prototypes that demonstrate connectivity, edge processing, and security — but your in-house teams are bottle-necked by long development cycles and unpredictable integration issues.
4. Agricultural & Rural Networking Providers
You need rugged, affordable devices to extend connectivity into fields, barns, and remote communities — but commercial gear is overpriced, hard to customize, and not designed for rapid prototyping or deployment in challenging environments.
5. Telecom & Network Operators
You need cost-effective, rapidly deployable edge devices for monitoring network performance, testing bandwidth in rural or urban environments, or validating new customer premises equipment—but traditional hardware procurement cycles are too slow and expensive.
6. Smart City & Infrastructure Providers
You’re deploying IoT devices to manage traffic lights, utilities, or environmental sensors across a city, but you need quick, low-cost prototypes to validate integrations before scaling to tens of thousands of units.
7. Educational & Research Institutions
Your students or researchers need reproducible, documented environments for experimentation with hardware, networking, or AI, but setting up reliable builds consumes valuable teaching and research time.
8. Healthcare & MedTech Device Innovators
You’re exploring connected health devices—remote patient monitors, smart diagnostic tools, or secure data collection endpoints—but you need a prototype that proves functionality while meeting strict reliability and security requirements.
9. Defense & Public Safety Contractors
You’re tasked with rapidly developing ruggedized, secure edge devices for field communication, surveillance, or sensor fusion, but your internal teams can’t keep pace with the prototyping demands.
10. Environmental & Energy Monitoring Firms
You need distributed, low-power devices to collect data in harsh or remote environments—forests, farms, offshore rigs, or mines—but your current prototypes fail due to durability or reproducibility issues.
11. Media & Event Production Companies
You want portable, reliable devices for live-streaming, crowd analytics, or on-site Wi-Fi provisioning at concerts and sporting events, but consumer gear isn’t flexible enough and enterprise hardware is overkill.
12. Transportation & Logistics Providers
You’re experimenting with fleet tracking, warehouse automation, or smart inventory systems, but you need a way to test edge hardware integrations quickly before committing to full-scale rollouts.
13. Industrial Automation & Robotics
You need controllers and monitoring systems for robots, conveyors, or factory IoT sensors, but the cost and time of custom PLCs and proprietary systems make it hard to experiment quickly.
14. Consultancies & Systems Integrators
You’re responsible for stitching together hardware and software for your clients, but you lack a streamlined way to spin up reproducible prototypes that demonstrate proof-of-concept value quickly and reliably.
