<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by ZElfeheil on Medium]]></title>
        <description><![CDATA[Stories by ZElfeheil on Medium]]></description>
        <link>https://medium.com/@zElfeheil?source=rss-1584bc8cacca------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*L42PEestuYxN8a1A</url>
            <title>Stories by ZElfeheil on Medium</title>
            <link>https://medium.com/@zElfeheil?source=rss-1584bc8cacca------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:19:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zElfeheil/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Porting Android Automotive OS (AAOS) to Raspberry Pi 4 as a QNX Guest VM: A Field Journal]]></title>
            <link>https://medium.com/@zElfeheil/android-automotive-to-raspberry-pi-4-as-a-qnx-guest-vm-4fb073a2c0d8?source=rss-1584bc8cacca------2</link>
            <guid isPermaLink="false">https://medium.com/p/4fb073a2c0d8</guid>
            <category><![CDATA[android-automotive]]></category>
            <category><![CDATA[raspberry-pi]]></category>
            <category><![CDATA[qnx]]></category>
            <category><![CDATA[hypervisors]]></category>
            <category><![CDATA[aao]]></category>
            <dc:creator><![CDATA[ZElfeheil]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 09:23:13 GMT</pubDate>
            <atom:updated>2026-07-06T09:29:55.115Z</atom:updated>
            <content:encoded><![CDATA[<p>If you’ve ever tried compiling Android Automotive OS (AAOS) for an embedded target, you know it’s a test of patience. Now, try taking that compiled image, ripping out its physical hardware dependencies, and forcing it to run as a virtualized Guest OS under a QNX Hypervisor on a humble Raspberry Pi 4.</p><p>It’s a journey filled with compilation roadblocks, memory-mapping puzzles, and deep dives into Android’s first-stage initialization (init).</p><p>After countless hours of debugging, I successfully bridged AAOS and QNX on the RPi4. This is the engineering journal of how I got there, the architectural choices I made, and the hidden traps you can avoid.</p><h3>The Architecture Strategy</h3><p>Before looking at the code, let’s look at the target landscape. We are partitioning an 8GB Raspberry Pi 4 to host both QNX and Android Automotive simultaneously:</p><h3>Phase 1: The Initial Struggle &amp; The Pivot</h3><p>Initially, I aimed for the stars: <strong>Android 14 (AOSP)</strong> built locally on Apple Silicon and via Docker on an Ubuntu 24.04 environment.</p><p>The environment broke immediately. AOSP toolchains fought bitterly with glibc/kernel version mismatches on Ubuntu 24.04, leading to silent signal: killed errors and abrupt exit status 2 failures.</p><h3>The Choices That Saved the Build:</h3><ol><li><strong>The Infrastructure Switch:</strong> I abandoned local virtualization and spun up a native <strong>Hetzner CPX62 server </strong>running Ubuntu 22.04 LTS (I had the built just in less than 1 hour, and it costed me around 2 €).</li><li><strong>The Software Strategic Pivot:</strong> Instead of fighting vanilla AOSP manifest mismatches, I opted for Chris Simmonds’ fantastic <strong>A3M (Android Automotive Architecture)</strong> project repository (I had some struggles with the latest AAOS, so I opted for stable version even if it is older).</li><li><strong>Beating the OOM Killer:</strong> Even on a robust cloud server, AOSP compilation is a memory hog. I provisioned a <strong>16GB Swap file</strong> and throttled parallel compilation down to -j16.</li></ol><p>After resolving a broken arm-trusted-firmware line in the manifest using a quick sed script, the build finally crossed the finish line, spitting out a clean, 8.0GB baseline sdcard.img.</p><h3>Phase 2: Preparing for the QNX Hypervisor (VirtIO Implementation)</h3><p>To transition Android from running on bare-metal Raspberry Pi hardware to running inside a QNX Virtual Machine (QVM), we had to completely swap its underlying driver model to <strong>VirtIO</strong>.</p><h3>1. Kernel Overhaul</h3><p>I modified the kernel configuration (rpi4_defconfig) to bake VirtIO support directly into the core compilation rather than as external modules:</p><p>Code snippet</p><pre>CONFIG_VIRTIO=y<br>CONFIG_VIRTIO_MMIO=y<br>CONFIG_VIRTIO_BLK=y<br>CONFIG_VIRTIO_NET=y<br>CONFIG_VIRTIO_GPU=y<br>CONFIG_VIRTIO_CONSOLE=y<br>CONFIG_VIRTIO_DMA_SHARED_BUFFER=y<br>CONFIG_BLK_DEV_DM=y<br>CONFIG_DM_LINEAR=y</pre><p><em>Crucial Realization:</em> The QNX Hypervisor implements VirtIO via <strong>Memory-Mapped IO (MMIO)</strong>, not PCI. Forgetting CONFIG_VIRTIO_MMIO will trigger an immediate kernel panic: VFS: Cannot open root device.</p><h3>2. Android Product Layer Changes</h3><p>Next, we told Android to drop the physical Broadcom graphics hardware and leverage virtualized HALs. In device/glodroid/rpi4/rpi4car.mk, I appended the VirtIO frameworks:</p><p>Makefile</p><pre>PRODUCT_PACKAGES += \<br>    android.hardware.graphics.composer@2.1-service.virtio \<br>    android.hardware.graphics.allocator@2.0-service.virtio \<br>    android.hardware.audio@6.0-impl.virtio</pre><h3>Phase 3: The Boot Barriers (And How to Smash Them)</h3><p>Compiling successfully didn’t mean it booted. The Android user-space security and partition architecture actively resist running in a virtual environment. Here are the critical barriers encountered during testing and how they were bypassed:</p><h3>1. The SELinux Crash</h3><p><strong>The Problem:</strong> First-stage init crashed instantly when attempting to mount selinuxfs with an Invalid argument.</p><p><strong>The Fix:</strong> The Broadcom baseline configuration had security modules disabled. I had to explicitly enforce CONFIG_SECURITY_SELINUX_DEVELOP=y and CONFIG_SECURITY_SELINUX_BOOTPARAM=y to let the system boot into permissive mode.</p><h3>2. Ghost Device Trees &amp; Dynamic Mapping</h3><p><strong>The Problem:</strong> First-stage init failed with realpath failed: /dev/block/by-name/super: No such file or directory. QNX merges the host&#39;s physical DTB into the guest, causing AOSP to aggressively look for physical hardware controllers instead of our virtualized block devices.</p><p><strong>The Fix:</strong> I passed androidboot.android_dt_dir=/nonexistent on the kernel command line. This forced init to ignore the corrupt Device Tree layout and fall back to the command line parameters where I manually declared the VirtIO controllers: androidboot.boot_devices=1c0d0000.disk,vdevs.</p><h3>3. Fstab Custom Ramdisk Injection</h3><p><strong>The Problem:</strong> Early mounting failed because Android couldn’t read the fstab out of the dynamically created QVM environment.</p><p><strong>The Fix:</strong> I unpacked the legacy LZ4 boot ramdisk, injected a hardcoded /fstab.rpi4car file mapping the system, vendor, and product partitions straight to /dev/block/vda, and repacked it.</p><h3>4. Deep Dive: Decoding the GPT Matrix to Target Disk Offsets</h3><p>To bypass deep user-space blocks like File-Based Encryption (FBE) and ext4 filesystem quotas on /data (userdata), I had to use Python patcher scripts to safely rewrite parameters right inside the raw binary image (sdcard_virtio.img).</p><p>To locate where these partitions started, the script parsed the image’s <strong>GUID Partition Table (GPT)</strong> header. In a standard GPT layout, partition names are allocated a maximum of 72 bytes and stored explicitly as <strong>Little-Endian UTF-16 (UTF-16LE)</strong> text.</p><p>For instance, when reading the table bytes to find the raw offset for Partition 1 (bootloader), the script scanned for this exact hex sequence:</p><p>62 00 6f 00 6f 00 74 00 6c 00 6f 00 61 00 64 00 65 00 72 00</p><p>Parsed letter-by-letter as little-endian words, it maps exactly to the ASCII equivalent:</p><ul><li>62 00 ➔ <strong>b</strong></li><li>6f 00 ➔ <strong>o</strong></li><li>6f 00 ➔ <strong>o</strong></li><li>74 00 ➔ <strong>t</strong></li><li>6c 00 ➔ <strong>l</strong></li><li>6f 00 ➔ <strong>o</strong></li><li>61 00 ➔ <strong>a</strong></li><li>64 00 ➔ <strong>d</strong></li><li>65 00 ➔ <strong>e</strong></li><li>72 00 ➔ <strong>r</strong></li></ul><p>By identifying these name blocks dynamically, the patcher found the strict structural anchor boundaries of the disk image:</p><ul><li><strong>Bootloader Partition (</strong><strong>Start LBA = 256 / </strong><strong>Hex 0x100):</strong> The sector offset where the primary Android bootloader code lives.</li><li><strong>U-Boot Environment (</strong><strong>Start LBA = 262144 / </strong><strong>Hex 0x40000):</strong> The sector address hosting the uboot-env variable block.</li><li><strong>Userdata Partition (</strong><strong>Start LBA = 4249616 / </strong><strong>Hex 0x40D810):</strong> This maps directly to byte offset 2175803392(0x81B02000) on the disk. Using this calculation, the python patcher (modify_userdata.py) targeted the exact sector bounds to toggle off the ext4 quota flags inside the superblock!</li></ul><h3>5. Low Memory Killer Daemon (lmkd) Loop</h3><p><strong>The Problem:</strong> Android would boot up to the 25-second mark and suddenly reset because lmkd failed continuously.</p><p><strong>The Fix:</strong> lmkd relies heavily on Kernel Pressure Stall Information (PSI). While the kernel compiled with CONFIG_PSI=y, it was disabled by default at boot time. Adding psi=1 to the kernel command line argument fixed it instantly.</p><h3>The Debugging Shortcut: Why I Compiled Locally</h3><p>During this grueling phase of trial-and-error, I hit a massive bottleneck. Every time a new boot barrier popped up, I needed to tweak a kernel config flag or modify an initialization setup. Running a full AOSP rebuild on the Hetzner server just to change a couple of flags was incredibly time-consuming and slowed down my debugging momentum.</p><p>To bypass this issue, I used a shortcut: <strong>I ran </strong><strong>rebuild_kernel.sh locally on my Mac.</strong> It was exponentially faster to cross-compile just the isolated kernel locally, package it, and swap it out, rather than triggering a massive, heavy cloud build every single time.</p><blockquote><strong><em>Note for Future Clean Builds:</em></strong><em> Once you clear the debugging phase, you can configure your AOSP source tree on the Hetzner server properly from the start. By embedding the VirtIO/PSI flags in the native kernel defconfig and pre-configuring the </em><em>fstab configurations to handle the unencrypted userdata and </em><em>slotselectnatively, you can run the build command on Hetzner once. It will compile everything cleanly, completely cutting out the need for local </em><em>rebuild_kernel.sh steps or post-build Python patcher scripts. The output </em><em>sdcard.img will be instantly ready for QNX!</em></blockquote><h3>The Reward: Hypervisor Memory &amp; Image Load Offsets</h3><p>With everything configured, the final architecture relies on an exact division of the Raspberry Pi 4’s 8GB physical space. We allocate <strong>4GB to Android</strong> starting at base address 0x80000000 (avoiding low-level peripheral overlaps), leaving a healthy 4GB behind for the QNX host to handle scheduling, device emulation, and page tables safely.</p><p>When deploying and handling image boots, paying absolute attention to memory layout addresses inside the QNX Makefile build parameters and android.qvmconf configuration is vital:</p><ul><li><strong>0x80000 (Decimal </strong><strong>524288):</strong> The definitive QNX host image load address explicitly targeted inside the main QNX Makefile.</li><li><strong>0x80080000:</strong> The Guest VM Linux Kernel load offset defined inside android.qvmconf. This maps cleanly to the Guest RAM base of 0x80000000 + the standard +512KB load padding expected by ARM64 Linux kernels to initialize code execution boundaries safely.</li></ul><p><strong>ComponentGuest Physical Address (GPA)Host Physical Address (HPA)DescriptionGuest RAM</strong>0x80000000 - 0x17FFFFFFFDynamic (MMU Stage 2)4GB RAM allocated to Android VM <strong>Kernel Load Address</strong>0x80080000Inside Guest RAMBaseline load offset (+512KB from base)<strong>VirtIO Block Device</strong>0x1C0D0000N/A (Software Emulated)The MMIO window for our sdcard_virtio.img</p><h3>How to Package This for the Community</h3><p>If you want to try this yourself or distribute it, there are two clear integration avenues:</p><ul><li><strong>The Fast Binary Deployment Path:</strong> Provide a pre-patched ramdisk, a clean android.qvmconf, and a Python automated script utility (patch_super_fstab.py) to post-process any generated GloDroid image over an SSH link.</li><li><strong>The Native AOSP Integration Path:</strong> Directly patch the source repository tree. Append the CONFIG_VIRTIOflags into device/glodroid/rpi4/kernel.config and adapt fstab.rpi4 natively to disable quotas and encryption right from the source build.</li></ul><h3>Moving Forward</h3><p>Seeing an Android Automotive stack successfully log via hvc0 virtual console while being managed by QNX on consumer-grade 170 € hardware is incredible. This validates the feasibility of testing mixed-criticality systems right on your desk without expensive automotive development boards.</p><p><em>Next up on the bench:</em> Mapping some signal into the QNX layer and routing it gracefully to the Android VHAL!</p><p><em>Have you worked with embedded hypervisors or Android OS compilation challenges? Let’s talk about it in the comments below!</em></p><h3>References &amp; Resources</h3><p>To help you get started or deep-dive further into the specific layers of this stack, here are the foundational resources used throughout this project:</p><ul><li><strong>Project GitHub Repository:</strong> <a href="https://www.google.com/search?q=https://github.com/ZElfeheil/qvm-android-rpi4">ZElfeheil/qvm-android-rpi4</a> (Contains my android.qvmconf, custom scripts, structural patchers, and step-by-step instructions to reproduce this build).</li><li><strong>QNX Hypervisor BSP &amp; Board Support:</strong> <a href="https://gitlab.com/qnx/hypervisor/getting-started?ref=devblog.qnx.com">QNX Hypervisor Getting Started GitLab</a> (Official starting guide and reference configurations for QVM virtualization).</li><li><strong>GloDroid Project Source:</strong> <a href="https://github.com/GloDroid">GitHub — GloDroid</a> (Core hardware integration layers for Android on Single Board Computers).</li><li><strong>A3M Android Automotive Manifest:</strong> <a href="https://github.com/aospandaaos/a3m-rpi-manifest">GitHub — a3m-rpi-manifest</a> (Chris Simmonds’ architecture manifest optimized for AAOS on RPi4).</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fb073a2c0d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Perception Studio: A Browser Tool for Early-Stage Sensor-Stack Design]]></title>
            <link>https://medium.com/@zElfeheil/perception-studio-a-browser-tool-for-early-stage-sensor-stack-design-90c5d660da82?source=rss-1584bc8cacca------2</link>
            <guid isPermaLink="false">https://medium.com/p/90c5d660da82</guid>
            <category><![CDATA[radar]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[perception]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[sensor-fusion]]></category>
            <dc:creator><![CDATA[ZElfeheil]]></dc:creator>
            <pubDate>Wed, 04 Mar 2026 15:16:07 GMT</pubDate>
            <atom:updated>2026-03-04T15:16:07.716Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3J5avNYiNobW9KyPw5JKTw.png" /></figure><h3>Perception architecture decisions tend to get locked in early, often before anyone has solid evidence that the sensor stack is actually “right”.</h3><p>In practice, a lot of projects still answer questions like these with rough assumptions:</p><ul><li>Where are the blind zones with this placement?</li><li>How much overlap do we really have for fusion?</li><li>What detection performance should we expect (especially for radar)?</li><li>How much compute are we implicitly committing to?</li></ul><p>If you’re unlucky, you only learn the answers after integration — when it’s expensive to change.</p><p>I built <strong>Perception Studio</strong> to make this early stage more explicit and easier to iterate on:</p><p><a href="https://perceptionstudio.net/">https://perceptionstudio.net/</a></p><p>It’s not meant to replace high-fidelity simulation. It’s a lightweight way to explore the system-level questions that usually sit between datasheets, rough sketches, and late integration surprises.</p><h3>The gap: tools for algorithms, fewer tools for architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q4uLet7rlwa7hHYrVHYlbQ.png" /></figure><p>Most perception tooling is built around:</p><ul><li>datasets and labeling</li><li>model training and evaluation</li><li>tracking and fusion algorithms</li><li>offline/online metrics</li></ul><p>But system architecture work is often fragmented:</p><ol><li>sensor placement sketches in slides or CAD</li><li>range assumptions pulled from datasheets</li><li>back-of-the-envelope radar coverage estimates</li><li>integration → field testing → rework</li></ol><p>Those early choices drive a lot of downstream complexity:</p><ul><li>cost and wiring/integration effort</li><li>redundancy and safety envelope</li><li>compute load and latency budgets</li></ul><h3>What Perception Studio does</h3><p>Perception Studio is a browser-based tool for early-stage perception architecture exploration. The focus is:</p><ul><li>placing sensors (radar / camera / LiDAR) in a 3D environment</li><li>visualizing coverage and blind spots</li><li>exploring trade-offs early, before hardware is fixed</li><li>thinking through environmental and compute implications</li></ul><p>The intent is to reduce “architecture by inertia” and make iteration faster.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-t33jfhS_IPBqQ5CfHvK8Q.png" /></figure><h3>Sensor layout iteration in 3D</h3><p>Sensor placement is one of the easiest things to underestimate.</p><p>Small placement changes can create:</p><ul><li>unexpected blind zones (near-field and oblique angles)</li><li>poor overlap (hurts fusion robustness)</li><li>calibration headaches</li><li>occlusion issues from platform geometry</li></ul><p>Being able to iterate layouts quickly helps narrow the search space early — even if you later validate the final design with a more detailed simulation stack.</p><h3>Environmental thinking early</h3><p>Perception failures are often environmental failures.</p><p>Even without full physics simulation, it’s useful to force the question early:</p><ul><li>Which sensor carries the system in degraded visibility?</li><li>What happens when camera performance collapses?</li><li>Where is LiDAR likely to drop out?</li><li>Is radar alone sufficient for certain detection tasks?</li></ul><p>The tool is designed to support that style of “architecture reasoning” early on.</p><h3>Compute and complexity awareness</h3><p>Every sensor you add increases pipeline complexity:</p><ul><li>bandwidth and synchronization</li><li>calibration maintenance</li><li>fusion graph complexity</li><li>inference and tracking load</li><li>latency constraints</li></ul><h3>Try it / feedback</h3><p>If you want to play with sensor-stack layouts and reason about coverage earlier in the process:</p><p><a href="https://perceptionstudio.net/">https://perceptionstudio.net/</a></p><p>If you have feedback on what would make it more useful (features, models, workflow), I’m happy to iterate.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90c5d660da82" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[BMAD-METHOD Brainstorming: A Structured Workflow for Creative Exploration]]></title>
            <link>https://medium.com/@zElfeheil/bmad-method-brainstorming-a-structured-workflow-for-creative-exploration-1ba0ff32ab51?source=rss-1584bc8cacca------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ba0ff32ab51</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[creative-thinking]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[brainstorming]]></category>
            <dc:creator><![CDATA[ZElfeheil]]></dc:creator>
            <pubDate>Tue, 17 Feb 2026 14:44:27 GMT</pubDate>
            <atom:updated>2026-02-17T14:44:27.323Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently, I attended a meetup where I came across the <a href="https://github.com/bmad-code-org/BMAD-METHOD"><strong>BMAD-Method</strong></a>. I wanted to understand it beyond the headline, so I cloned the repository and experimented with it using <strong>Google’s Antigravity</strong> as the interface for running workflows and agents. I began by asking Antigravity to explain the framework and how to use it.</p><p>At a high level, BMAD positions itself as a <strong>structured, </strong>Agile Ai AI-driven development aligned with the Agile software development lifecycle. The idea is not to replace the lifecycle, but to make each step explicit, guided, and easier to document across stakeholders.</p><h3>The “standard journey” (end-to-end)</h3><p>BMAD can be understood as a full path from an initial idea to a running product, moving through planning, design, implementation, and feedback loops.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GFjhUEDJBc2LqbMQoIGsbQ.jpeg" /><figcaption>Fig 1. BMAD standard journey</figcaption></figure><p>However, this article is not about the full pipeline. What stood out to me most was a <strong>utility workflow</strong> bundled inside BMAD: <strong>brainstorming</strong>.</p><h3>The brainstorming utility: a 4-step workflow</h3><p>The BMAD brainstorming workflow is short, but it is not “lightweight” in the casual sense. It is designed to protect creative exploration while still producing a usable output.</p><p>The flow is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t-HBNZbtzXKRBTF5Xix4gw.jpeg" /><figcaption>Fig 2. BMAD brainstorming</figcaption></figure><p>Before starting, I asked Antigravity to explain what this workflow actually does. This is the essence of what it returned:</p><h3>1. Session Setup &amp; Discovery</h3><ul><li>I’ll act as your Facilitator.</li><li>I start by asking two key questions to ground us:</li></ul><p>What are we brainstorming about? (Topic/Challenge), and What specific outcomes are you hoping for? (Goals/Insights)</p><ul><li>Once I understand the context, I initialize a session document:</li></ul><p>_bmad-output/brainstorming/brainstorming-session-{{date}}.md</p><h3>2. Choose Your Path</h3><p>You’ll interpret your needs and pick one of four modes:</p><ul><li>[1] User-Selected: You browse the library of ~60 techniques (like <em>SCAMPER</em>, <em>Six Thinking Hats</em>, <em>Reverse Brainstorming</em>) and pick the ones you want.</li><li>[2] AI-Recommended: I analyze your goals and suggest the best 3–4 techniques to get there.</li><li>[3] Random Selection: We roll the dice! Good for breaking out of ruts.</li><li>[4] Progressive Flow: A structured journey starting broad (Divergent) and narrowing down (Convergent).</li></ul><h3>3. The Brainstorming Loop</h3><ul><li>We execute one technique at a time.</li><li>Anti-Bias Protocol: Every 10 ideas, I will force a pivot (e.g., from “Technical” to “User Experience” to “Business Viability”) to prevent us from getting stuck on one track.</li><li>Goal: We aim for 100+ ideas. The first 20 are usually obvious; the real gold is often in ideas 50–100.</li></ul><h3>4. Organization &amp; Next Steps</h3><ul><li>Once we have a pile of ideas, we switch to Convergence Mode.</li><li>We group, filter, and prioritize them.</li><li>Finally, we can output them into a format ready for the Product Brief or PRD workflows.</li></ul><p>What I liked most about the BMAD workflow was the explicit use of creativity techniques in step 2. Techniques function as “engines”: they trigger different cognitive moves rather than relying solely on free-form ideation.</p><p>BMAD groups techniques into 8 categories, such as:</p><ol><li>Collaborative 🤝 — e.g., Yes-And building, Brainwriting, Random stimulation</li><li>Creative 🎨 — e.g., What-if scenarios, Analogies, Reversal, First principles</li><li>Deep 🧠 — e.g., Five Whys, Assumption reversal, Morphological analysis</li><li>Introspective ✨ — e.g., Future self interview, Values archaeology</li><li>Structured 📐 — e.g., SCAMPER, Six Thinking Hats, Mind mapping</li><li>Theatrical 🎭 — e.g., Alien anthropologist, Persona journey</li><li>Wild 🦁 — e.g., Anti-solution, Zombie apocalypse planning</li><li>Biomimetic 🌿 — e.g., Nature’s solutions, Ecosystem thinking</li></ol><p>Even though I used the <strong>AI-recommended</strong> path, I reviewed the library to see what I could select manually. Some techniques were familiar (for example, <strong>Six Thinking Hats</strong>), while many were new to me.</p><p><strong>Note: Six Thinking Hats</strong> was introduced by <strong>Edward de Bono</strong> in his 1985 book <em>Six Thinking Hats</em>.</p><h3>A strong extension: market research → brainstorming</h3><p>One extension I found particularly practical was combining brainstorming with structured market research. <strong>BMAD </strong>supports a workflow in which a market research step (in the analysis phase) produces a structured document that can then be used as input for brainstorming.</p><p>In my case, this mattered because I already had a defined goal: I wanted the brainstorming session to be anchored in <strong>competitor gaps</strong> and <strong>customer pain points</strong>, rather than abstract ideation.</p><h3>My takeaway</h3><p>For me, BMAD brainstorming is not just chatting for ideas like ChatGPT or others. It is a structured technique to formulate or discover ideas.</p><p>That last part (the delivered artifact) is important. It improves traceability and makes downstream execution more coherent because your planning deliverables are explicitly linked to a documented exploration process.</p><p>If you work in <strong>innovation management</strong> or <strong>product development</strong>, it is worth trying at least once, especially if you want brainstorming to end in something more actionable than a list of ideas.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1ba0ff32ab51" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>