AI Engineer

500 people vibe-coded for 30 days. I was one of them. - Sanja Grbic, Automattic

4964 summary words 22 min summary Watch video

Start with the signal

22 min read

Summary

At-a-Glance

  • Verdict: Watch fully
  • Core thesis: AI-assisted coding tools combined with organizational permission structures can transform non-engineers into shipping contributors in 30 days, but the unlock requires systematic enablement (training, security setup, documentation access) plus the cultural space to experiment without roadmap pressure.
  • Why it matters: Provides concrete evidence of role transformation velocity at scale (500 people, 794 projects) in a 1,400-person distributed company, showing both what shipped and how organizational design enabled it—directly applicable to your operator/agent workflow questions.
  • Best use: Study for implementing AI enablement programs and understanding the infrastructure prerequisites (MCP servers for context, dev environment docs, security processes) that gate non-technical contributors from actual production shipping.

Executive Summary

Sanja Grbic, a product designer at Automattic (WordPress.com, Jetpack, WooCommerce, Tumblr, 1,400 employees, fully distributed), describes "Radical Speed Month"—a company-wide experiment where ~500 people paused roadmaps for 30 days to pair-ship real projects. She went from reviewing PRs to shipping three production projects in 30 days: a board game session manager (2 hours, group of 4), a design system status tracker (2.5 weeks solo), and a WooCommerce iOS chat with AI agent (6 days with one other designer). The speed unlocks came from three organizational preconditions: mandatory two-week immersive AI enablement courses tailored by role, systems/ops setup of security/docs good enough for non-engineers to spin up dev environments, and an MCP server (Context AHC) giving AI tools access to decades of internal documentation.

The talk is not about what AI can do in isolation—it's about the organizational design choices that let non-engineers ship to production in legacy systems. Automattic systematically enabled this: access to tools/courses, role-specific AI training, encouragement to contribute code, and crucially, the 30-day permission structure to pause roadmaps and own full stack without handoff/negotiation. Sanja's process shifted from Figma-first to code-prototype-first, with visual tools used only for late-stage fine-tuning. The engineer's role evolved from builder to enabler/teacher, setting up Git, explaining workflows, and unblocking non-technical contributors.

Key tactical detail: the third project (iOS chat) used a "Cloud Code project folder" where all chats/ideas were recorded into the file system, which "significantly sped up collaboration" and became her ongoing practice. This is a concrete example of using LLM conversation logs as structured project artifacts. The design system tracker pulled live data from GitHub, Storybook, and Figma; sorted by status; tagged by library; with working search—built solo by a designer in 2.5 weeks. The WooCommerce chat includes WordPress.com auth, Jetpack connection, site-themed widget, and an AI agent that scans sites and decides answerable vs. not—shipped in 6 days by two designers.

The meta-lesson: "Changing processes or tools in large organizations requires shifting the human behavior behind them." Automattic's approach was access + enablers/champions + experimentation space + agency. The 794 projects from 500 people in one month is the scale proof. For Ken: this is a playbook for enabling non-traditional contributors in agent/AI workflows, showing that the bottleneck is not AI capability but org design (training cadence, security/docs, context servers, permission to experiment). The speed came from removing negotiation/handoff, not just from the tools.

Key Takeaways

  • Claim: Automattic ran "Radical Speed Month": ~500 employees (1/3 of company) paused roadmaps for 30 days, paired up, and shipped 794 projects with full autonomy. | Evidence: Sanja built three projects: 2-hour board game manager (group of 4), 2.5-week design system status tracker (solo), 6-day WooCommerce iOS chat with AI agent (pair). Instructions were: pause roadmap, work 30 days, pair with partner, ship something real, AI optional. | Caveat: Required staging across company; not all 1,400 at once. Success depended on pre-existing enablement infrastructure (training, security, docs, context server). No mention of how many of 794 projects reached production or sustained use. | Implication: Large orgs can run month-long AI experimentation sprints if they front-load enablement and grant permission to pause roadmaps. The 794-project output is a forcing function for culture change, not just skill development. For Ken's agent systems, this model could validate operator workflows at scale. | Timestamp: timestamp unavailable
  • Claim: Non-engineers can ship to production in legacy systems if given systematic enablement: mandatory 2-week immersive AI training per role, ops-setup security/processes, and MCP server access to internal docs. | Evidence: Sanja had "some knowledge of GitHub, reviewed PRs, played with Claude Code" but never shipped code at work. After 2-week AI enablement course + Context AHC (MCP server for internal knowledge) + ops team's dev environment docs, she deployed design system tracker to internal platform and external site, plus iOS app with WordPress.com auth and Jetpack connection. | Caveat: Automattic is fully distributed/asynchronous with "decades" of well-documented work, which Context AHC surfaces. Companies without mature docs or MCP-style context retrieval would face friction. No specifics on security review process or post-experiment maintenance handoff. | Implication: Building MCP servers for internal context is a prerequisite for enabling non-technical contributors at scale. For Ken: agent systems need equivalent context retrieval to unblock operators; the training+docs+security combo is the real unlock, not just tool access. | Timestamp: timestamp unavailable
  • Claim: The engineer's role shifts from builder to enabler/teacher when non-engineers start coding; this may have "far greater impact" than doing more engineering oneself. | Evidence: In the 2-hour board game manager project, the engineer "set up the project in GitHub, helped us grasp basic commands, helped us understand how Git works." Sanja notes she "thought I understood Git, but actually I didn't—I really needed someone to tell me how it works and what are the best ways to set up a project." All four team members (two designers, one engineer, one product lead) committed code. | Caveat: This role shift only matters if the org creates space for non-engineers to contribute; otherwise it's wasted effort. Sanja had no-risk environment (project unrelated to existing products) focused on learning. No discussion of how this scales when non-engineers ship to production systems with real user impact. | Implication: For Ken's operator/agent workflows: design roles where senior engineers are measured on enablement, not just output. The 2-hour collaborative build unlocked her later solo 2.5-week project. Enablers are force multipliers in AI-assisted environments. | Timestamp: timestamp unavailable
  • Claim: Process inversion: the main artifact became the code prototype, not Figma mocks; Figma was used only for late-stage visual fine-tuning and mood boards. | Evidence: Design system tracker: "built the prototype initially, then used Figma only later for some visual fine-tuning, some things I just couldn't explain to Claude Code." iOS chat: "we had a very good plan, then we built the prototype/product, then went back into Figma... to build a mood board and for some UI fine-tuning." Prototype built in 1 week, deployment took another 1.5 weeks. | Caveat: This works when AI can generate functional UIs quickly and the designer can iterate in code. If the codebase is brittle or the designer lacks env setup, the handoff still bottlenecks. No mention of design system constraints or accessibility review. | Implication: Traditional design handoff (high-fidelity mocks → eng build) is bypassed when designers code. For Ken: agent-generated UIs could skip Figma entirely if operators can tweak in-place. The 1-week prototype vs. 1.5-week deployment split suggests infra/security remains the slow part. | Timestamp: timestamp unavailable
  • Claim: Using a "Cloud Code project folder" where all chats and ideas are recorded into the file system significantly speeds collaboration and build; this became her ongoing practice. | Evidence: For the iOS chat project (6 days, two designers), "we started the exploration as a Cloud Code project folder, where all of the chats and ideas were recorded into a file system. I noticed that significantly sped up collaboration, as well as the build later. And I've continued this practice in my daily work." | Caveat: No detail on folder structure, file format, or how the chats were referenced during build. Unclear if this is simple markdown logs or something more structured. No mention of tooling to query/surface past chats. | Implication: Treating LLM conversation logs as structured project artifacts (not ephemeral chats) is a tactical unlock. For Ken: agent systems should persist and version conversation history as first-class data, enabling later retrieval and team collaboration. This is adjacent to your MCP/context-passing work. | Timestamp: timestamp unavailable
  • Claim: The design system status tracker pulls live data from GitHub, Storybook, and Figma; sorts by status; tags by library; and includes working search—built solo by a designer in 2.5 weeks. | Evidence: "The tracker was picking up links from the GitHub repo, from Storybook, from Figma, sorting them according to status, tagging them with the correct library names. Search worked perfectly." Deployed internally and externally (curated data). Engineer initially questioned performance, maintenance, and feasibility. | Caveat: Proof of concept; no discussion of ongoing maintenance, data sync reliability, or performance at scale. External site had "curated data," suggesting manual curation or subset. No mention of API rate limits or auth complexity for Figma/GitHub/Storybook. | Implication: Non-engineers can build cross-platform integration dashboards in weeks if the APIs are accessible and the org allows solo ownership. For Ken: similar internal tools (surfacing agent logs, context usage, tool calls) could be built by operators, not just eng. The "no negotiation/handoff" speed is the key unlock. | Timestamp: timestamp unavailable
  • Claim: The WooCommerce iOS chat shipped in 6 days by two designers includes WordPress.com auth, Jetpack connection, site-themed widget, and a "fully capable AI agent" that scans sites, answers visitor questions in real-time, and decides answerable vs. not. | Evidence: "Fully working proof of concept that includes authentication through WordPress.com, includes a Jetpack connection... widget that inherits the site theming, where the site visitors can ask questions, as well as a fully capable AI agent that scans the site for information and answers visitors' questions in real time, and also discerns whether it can answer a question or not." | Caveat: Proof of concept; no mention of load testing, moderation, hallucination handling, or production readiness. "Scans the site for information" is vague—likely RAG over site content, but no architecture detail. No discussion of how the agent decides "answerable vs. not." | Implication: Two designers can ship agentic e-commerce features in a week when auth/infra are already in place. For Ken: this is evidence that agent-powered user-facing features are designer-shippable, not just eng-only. The 6-day timeline suggests speed came from reusing existing auth/connection primitives, not building from scratch. | Timestamp: timestamp unavailable

Detailed Brief

Organizational Enablement Infrastructure at Automattic

  • Claims: Automattic (1,400 employees, fully distributed/asynchronous) systematically prepared for AI velocity: access to tools/courses, then mandatory 2-week immersive AI enablement per role, encouragement to contribute code, ops-setup security/docs, and Context AHC (MCP server for internal knowledge).; Context AHC is an MCP server giving AI tools access to decades of documented work, used daily for research/planning.; Operations/systems team set up "fantastic security and processes as well as documentation that's good enough that a non-engineer such as myself can spin up a working development environment."
  • Evidence: Sanja: "We were encouraged to use it, to learn it, and to experiment with AI from the very beginning. At first we had access to tools and courses. Then the AI enablement project was initiated where every single employee goes through a two-week course designed specifically for their role. This is completely immersive with lectures and focused hands-on time."; "Over time as AI coding tools got better, everyone was encouraged to get comfortable with contributing in code."; "An incredibly helpful tool that I use every day for research and planning is Context AHC. It's an MCP server that gives AI tools access to all of the knowledge and data that we have documented over the years."; Automattic is "fully distributed and asynchronous which means that most of our work, our thoughts, our concepts from the past couple of decades are very well documented."
  • Caveats: No detail on what the 2-week AI enablement course contains beyond "lectures and focused hands-on time."; No specifics on security review process, PR approval workflows, or how non-engineers' code is vetted before production.; Context AHC (MCP server) effectiveness depends on mature internal documentation; companies with poor docs would not benefit equally.; No discussion of cost: training 1,400 employees for 2 weeks each is significant investment.
  • Implications: For Ken: MCP servers (or equivalent context retrieval) are foundational for enabling non-technical contributors. The 2-week training cadence is a heavy lift but appears necessary for large-org adoption.; The "encouragement to contribute code" + ops-setup docs is the difference between toy experiments and production shipping. Automattic front-loaded the enablement cost to unlock distributed velocity.; This infrastructure is what makes the 794 projects in 30 days possible. Without it, Radical Speed Month would have been a chaos of blocked contributors.; For agent/operator systems: equivalent context servers + role-specific training + dev-env-in-a-box are prerequisites for non-eng operators to ship agentic workflows.

Radical Speed Month Experiment Design and Results

  • Claims: ~500 people (1/3 of Automattic) participated in first round, starting ~794 projects in 30 days.; Instructions: pause the roadmap, work for 30 days, pair with a partner, ship something real. AI optional.; Sanja built three projects: 2-hour board game session manager (group of 4), 2.5-week design system status tracker (solo), 6-day WooCommerce iOS chat (pair with another designer).; The experiment was staged; not all 1,400 employees participated simultaneously.
  • Evidence: "About two months ago, we kicked off this project. The instructions were to pause the roadmap, work for 30 days, pair with a partner, and ship something real. We were welcome to use AI, though it wasn't a requirement."; "We had about a third of the company participating in the first round, that's around 500 people, that started around 794 projects in these 30 days."; Sanja's timeline: overlapped with 2-week AI enablement course, built three projects, each demonstrating increasing speed and autonomy.
  • Caveats: No data on how many of 794 projects reached production, sustained use, or were abandoned post-experiment.; No mention of coordination overhead, conflicts with existing roadmaps, or how teams handled the pause.; "AI optional" means results are a mix of AI-assisted and traditional workflows; unclear what % used AI heavily.; Staging across the company suggests logistical complexity; no discussion of how roles were assigned or teams formed.
  • Implications: 794 projects from 500 people in 30 days = ~1.6 projects per person on average, but Sanja's 3 projects suggest wide variance.; For Ken: month-long experimentation sprints with roadmap pauses are viable at scale if the org can absorb the coordination cost. The "pair with a partner" constraint likely improved shipping rate vs. solo work.; The experiment is a forcing function for culture change, not just a feature release mechanism. The goal is to shift "human behavior" and establish new processes.; For agent systems: similar 30-day operator sprints could validate agentic workflows and surface which roles can ship autonomously with AI assistance.

Project 1: Board Game Session Manager (Role Transformation)

  • Claims: Built in 2 hours by a group of 4 (two designers, one engineer, one product lead) during AI enablement training.; App allows creating/viewing/joining/managing board game sessions, with 16-bit office illustration, start/stop sessions, game/spot availability, and chat.; Engineer acted as enabler: set up GitHub project, taught basic Git commands, explained workflows. All four members committed code.; Biggest insight: "If you're an engineer, the impact that you have when you enable others may be far greater than the impact of doing more engineering yourself."
  • Evidence: "We decided to build this board game session manager. We had a board game evening planned that night. So we wanted to create an app that's used to create, view and join and manage board game event sessions."; "The engineer that was in our group was really great at setting us up for success. He set up the project in GitHub, helped us grasp some basic commands and helped us understand how Git works."; "We were all committing code to this project. And since this app that we were building wasn't related to any of our existing products, there were no risks involved and our focus was on learning."; Only tool: Claude Code. Visual tool: Nano Banana for office illustration.
  • Caveats: 2-hour build in a no-risk learning environment is not representative of production work.; No mention of app quality, bugs, or whether it was actually used for the board game evening.; "I thought I understood Git, but actually I didn't"—suggests even experienced designers lack foundational versioning knowledge.; The engineer's enablement role only matters if the org creates ongoing space for non-engineers to contribute; otherwise it's a one-time exercise.
  • Implications: For Ken: the 2-hour collaborative build is a microcosm of the enabler role. Senior engineers measured on teaching/unblocking (not just PRs merged) could be force multipliers in AI-assisted orgs.; The no-risk sandbox (unrelated to existing products) was crucial for psychological safety and rapid iteration.; All four roles contributing code suggests AI lowers the floor for technical contribution, but the engineer's setup/teaching was the critical dependency.; For agent workflows: operators need equivalent "setup and teach Git" moments to ship agentic systems; the first 2 hours of enablement unlock the next 30 days.

Project 2: Design System Status Tracker (Solo Build, Process Shift)

  • Claims: Built solo in 2.5 weeks by Sanja (designer) after engineer raised concerns about performance/maintenance/feasibility.; Pulls live data from GitHub, Storybook, Figma; sorts by status; tags by library; includes working search and live component previews.; Deployed internally and externally (external version had curated data).; Prototype built in 1 week, deployment/fixes took another 1.5 weeks.; Process shift: built prototype first, used Figma only later for visual fine-tuning; "the live project was what I was iterating on directly."
  • Evidence: "What I proposed, this design system status tracker, the solution that I proposed raised questions with the engineer around performance, around maintenance, and if we would be able to build it in the first place. But I was empowered by the previous exercise. I was allowed to experiment. So I decided to build the full idea as a proof of concept on my own."; "The tracker was picking up links from the GitHub repo, from Storybook, from Figma, sorting them according to status, tagging them with the correct library names. Search worked perfectly."; "I built the prototype initially as a start. And then I used Figma only later in the process for some visual fine tuning."; Delivered on internal platform and external public-facing site.
  • Caveats: Proof of concept; no discussion of ongoing maintenance burden or who owns it post-experiment.; External site used "curated data," suggesting manual curation or subset—not full automation.; Engineer's concerns (performance, maintenance) were not addressed in the talk; unclear if they materialized.; No detail on API rate limits, auth complexity, or data sync reliability for GitHub/Figma/Storybook.
  • Implications: For Ken: non-engineers can build cross-platform integration dashboards in weeks if APIs are accessible and the org grants solo ownership. The "no negotiation/handoff" speed is the unlock.; The 1-week prototype vs. 1.5-week deployment split suggests infrastructure/security/hosting is still the slow part, even with AI-assisted coding.; Process inversion (code-first, Figma-last) is viable when the designer can iterate in code; this is a workflow shift, not just a tool shift.; For agent systems: similar internal dashboards (agent logs, context usage, tool calls) could be operator-built, not eng-gated, if the org allows solo proof-of-concept ownership.

Project 3: WooCommerce iOS Chat with AI Agent (Speed Unlock)

  • Claims: Built in 6 days by two designers (Sanja + one other) after she had already adjusted her process from projects 1 and 2.; Fully working proof of concept: WordPress.com auth, Jetpack connection, site-themed widget for visitors, AI agent that scans site content and answers questions in real-time, decides answerable vs. not.; Alignment/ideation easy for two designers; "we always have the users in mind."; Started exploration as a "Cloud Code project folder" where all chats/ideas recorded into file system; "significantly sped up collaboration, as well as the build later."; Main artifact was prototype; Figma used for mood board and UI fine-tuning after prototype was built.
  • Evidence: "In only six days, we went from zero through the entire process to building an iOS chat for WooCommerce merchants that allows them to answer shoppers in real time, reply from their phone, or even set up AI to answer the questions for them."; "Fully working proof of concept that includes authentication through WordPress.com, includes a Jetpack connection... widget that inherits the site theming... fully capable AI agent that scans the site for information and answers visitors' questions in real time, and also discerns whether it can answer a question or not."; "We started the exploration as a Cloud Code project folder, where all of the chats and ideas were recorded into a file system. And I noticed that significantly sped up collaboration, as well as the build later. And I've continued this practice in my daily work."
  • Caveats: Proof of concept; no mention of load testing, moderation, hallucination handling, or production readiness.; "Scans the site for information" is vague—likely RAG over site content, but no architecture/embedding/retrieval detail.; "Discerns whether it can answer a question or not" is underspecified; no discussion of prompt engineering, guardrails, or failure modes.; 6-day timeline assumes WordPress.com auth and Jetpack connection primitives already exist; not building from scratch.
  • Implications: For Ken: two designers can ship agentic e-commerce features in a week when auth/infra are reusable. This is evidence that agent-powered user-facing features are designer-shippable, not eng-only.; The "Cloud Code project folder" practice (chats/ideas as file system artifacts) is a tactical unlock for collaboration. For agent systems, this suggests persisting conversation logs as structured project data, not ephemeral chats.; Speed came from both designers having "leveled up together and had a shared understanding on how we can apply the tools." The first two projects were training for the third.; For Ken's operator workflows: similar 6-day sprints could validate agentic features if operators have pre-built auth/context primitives and a shared process vocabulary.

Meta-Lessons on Organizational Change

  • Claims: First project (2 hours, group of 4): engineer's role elevated from builder to enabler; this skill "will be crucial as more and more people in other roles start working in code."; Second project (2.5 weeks, solo): moved Sanja from designer to design engineer, able to "push code to production in a well-established large system that requires serious onboarding and security practices."; Third project (6 days, pair): proof that speed was unlocked because both designers had adjusted process and "had a shared understanding on how we can apply the tools."; "In only 30 days, my process that has been similar for years has completely shifted. And that was true for many other people."; Key org design principle: "Changing processes or tools in large organizations requires shifting the human behavior behind them. So provide your people with access to the new tools, find your enablers and champions, create space for experimentation and give them agency so that they can break out of their habits and unlock the speed of the tools."
  • Evidence: "The first project showed me how the engineer's role can be elevated from a builder to an enabler. And this is a skill that I believe will be crucial."; "The second one moved me from a designer to a design engineer. And here I'm talking about being able to push code to production in a well-established large system."; "The third project... was proof that we've unlocked speed because I already had adjusted my process a little bit, but then we both leveled up together."; "All of the company initiatives that I described together gave me the space I needed to experiment and find my own process that works within what we already have established in the company."
  • Caveats: 30-day personal transformation is not the same as sustained behavior change; no follow-up data on whether Sanja continued shipping code post-experiment.; "Many other people" had similar shifts, but no quantitative data on how many, or what % of 500 participants.; Agency and experimentation space are easy to grant in a month-long experiment; harder to sustain when roadmaps resume.; No discussion of failure cases, projects that didn't ship, or people who didn't adopt AI workflows.
  • Implications: For Ken: the 30-day arc (enabler → solo build → collaborative speed unlock) is a pattern for role transformation. First project teaches the basics, second builds confidence/ownership, third proves the new process works.; The org design principle (access + enablers + space + agency) is the formula for large-org AI adoption. Tools alone don't shift behavior; the permission structure does.; For agent/operator systems: similar 3-project arcs could onboard operators to agentic workflows. First project: supervised agent build. Second: solo agent deployed. Third: collaborative multi-agent system.; The "break out of their habits" framing suggests the main barrier is psychological/cultural, not technical. Radical Speed Month is a forcing function to overcome inertia.

Notable Concepts & Terms

  • Radical Speed Month: Automattic's 30-day experiment where ~500 employees paused roadmaps, paired up, and shipped 794 projects with full autonomy. AI optional. Staged across company, not all at once. Goal: shift human behavior and establish new AI-assisted processes at scale.
  • Context AHC (MCP server): An MCP (Model Context Protocol) server Automattic built that gives AI tools access to decades of internal documentation (work, thoughts, concepts). Sanja uses it daily for research/planning. This is the context retrieval infrastructure enabling non-engineers to navigate legacy systems.
  • AI Enablement Project: Automattic's mandatory 2-week immersive training for every employee, tailored by role, with lectures and hands-on time. Precedes Radical Speed Month. Example of systematic preparation before allowing mass experimentation.
  • Cloud Code project folder: Sanja's practice (originated in project 3) of starting exploration in a folder where "all of the chats and ideas were recorded into a file system." This significantly sped up collaboration and build. Suggests treating LLM conversation logs as structured project artifacts, not ephemeral.
  • Engineer as enabler: Role shift where engineer's impact comes from setting up projects (GitHub, Git workflows) and teaching/unblocking non-engineers, rather than writing more code themselves. Sanja's claim: this "may be far greater than the impact of doing more engineering yourself" as more roles start coding.
  • Designer to design engineer transformation: Sanja's 30-day journey from reviewing PRs to pushing code to production in a large legacy system. Enabled by AI + org infrastructure (training, docs, security setup, MCP server) + permission to own full stack without handoff. Represents new hybrid role emerging from AI tooling.
  • Prototype-first process: Inversion of traditional design workflow: instead of Figma mocks → eng build, now build code prototype first, use Figma only for late-stage visual fine-tuning and mood boards. Main artifact is the live/working prototype, not the design file.
  • WordPress.com / Jetpack / WooCommerce ecosystem: Automattic's product suite. Jetpack is a WordPress plugin for security/performance/etc. WooCommerce is e-commerce platform. Sanja's projects integrated with these (auth, connections, merchant chat). Context: 20-year-old open-source ecosystem, 1,400 employees, fully distributed.

Operator Notes / Why Ken Should Care

  • MCP servers (Context AHC) are critical infrastructure for enabling non-technical contributors to navigate legacy systems. For agent systems: equivalent context retrieval is a prerequisite for operators to ship agentic workflows without eng handoff.
  • The 2-week AI enablement training (immersive, role-specific) is a heavy but necessary investment for large-org AI adoption. For Ken's workflows: similar onboarding programs for operators could accelerate agent/AI ops deployment.
  • Treating LLM conversation logs as structured project artifacts (Cloud Code project folder) is a tactical unlock for collaboration. For agent systems: persist/version conversation history as first-class data, enable later retrieval and team sharing.
  • The 30-day experiment arc (enabler project → solo build → collaborative speed unlock) is a pattern for role transformation. For agent ops: similar 3-project onboarding (supervised → solo → collaborative) could train operators to ship agentic systems.
  • The org design principle (access + enablers/champions + experimentation space + agency) is the formula for large-org AI adoption. For Ken: tools alone don't shift behavior; the permission structure (pause roadmaps, own full stack, no negotiation/handoff) does.
  • The 794 projects from 500 people in 30 days is scale proof that month-long experimentation sprints work if the org can absorb coordination cost. For Ken's GTM/investing: look for companies with similar enablement infrastructure (training, context servers, security docs) as signals of AI-native ops capability.
  • Engineer-as-enabler is a force multiplier role in AI-assisted orgs. For agent systems: design roles where senior engineers/operators are measured on teaching/unblocking, not just output. The first 2 hours of enablement unlock the next 30 days of velocity.
  • Prototype-first (code → Figma for fine-tuning) inverts traditional handoff. For agent UIs: operators could skip design tools entirely if they can tweak in-place. The 1-week prototype vs. 1.5-week deployment split suggests infra/security remains the slow part.
  • Two designers shipping an iOS chat with AI agent (auth, RAG, real-time answers) in 6 days is evidence that agentic user-facing features are designer-shippable when auth/context primitives are reusable. For Ken: similar speed unlocks possible in agent content/business workflows if primitives (context retrieval, tool calling, auth) are pre-built.
  • The design system status tracker (GitHub/Storybook/Figma integration, search, live previews, solo build in 2.5 weeks) shows non-engineers can build cross-platform dashboards if APIs are accessible and the org allows solo ownership. For Ken: internal agent monitoring/logging tools could be operator-built, not eng-gated.

Watch Map

  • timestamp unavailable: Transcript does not provide chapter timestamps or time markers. Video is 17:48 long. Content flows chronologically: intro/context → Automattic's AI enablement setup → Radical Speed Month overview → Project 1 (2-hour board game manager) → Project 2 (2.5-week design system tracker) → Project 3 (6-day WooCommerce iOS chat) → meta-lessons on org change.

Source/Metadata

  • Title: 500 people vibe-coded for 30 days. I was one of them. - Sanja Grbic, Automattic
  • Transcript words: 5052
  • Duration seconds: 1068
  • Timestamp note: Timestamps/chapters not present in transcript; video duration 17:48 (1068 seconds).
Full transcript 2548 words · 22 min read
0:00

SPEAKER_00

Hi there, welcome to my talk. My name is Sania and today I'm going to share with you an experiment we did at our company that's somewhat unusual and very exciting. It was a one-month initiative called Radical Speed Month where two-person teams were given full autonomy to build and ship projects. I built three projects during this time. I will walk you through them and we'll talk about the outcomes and learnings for myself as a designer as well as the system-wide impact this experiment had. I'll tell you a little bit about Automatic first. We're the company behind several products in the WordPress ecosystem such as WordPress.com, Jetpack, WooCommerce as well as many other brands like Tumblr, Beeper and many more. We are around 1400 people and we're fully distributed which is pretty interesting for also what I'm about to show you. We're fully distributed and asynchronous which means that most of our work, our thoughts, our concepts from the past couple of decades are very well documented. I'm a product designer and I've spent over a decade building software and watching how teams build software. I've worked at Automatic for five years now on the Jetpack design team and in the past I've worked for several companies of different sizes, different types of products and different team organizational configurations. So after all these years I've not only accumulated experience in designing products, I also have a lot of insight into team dynamics and organizational structures within product teams. Over the years the designer role has morphed and evolved a lot. We still have all of these different types of roles like web designers or UX designers or interaction or product designers. We've also used many different tools from the Adobe Suite to Sketch to Figma and many others in addition to those. And AI is a new tool. It is definitely not built just for designers. It's versatile so it can be used by many different roles but the expectation is the same for everyone. It's that it will bring us speed and of course this is why companies are rushing to put it to use. But no matter what role uses AI, this velocity is going to look very different in small teams like one to three people that are building new products, let's say a zero to one product versus large organizations that might have built entire product ecosystems that have software that's anywhere between five and twenty years old and that have more than a thousand or even thousands of people. Before I dive into the experiment, I will tell you a little bit about what preceded it in terms of how we use AI at Automatic. We were encouraged to use it, to learn it, and to experiment with AI from the very beginning. At first we had access to tools and courses. Then the AI enablement project was initiated where every single employee goes through a two-week course designed specifically for their role. This is completely immersive with lectures and focused hands-on time, very helpful. And over time as AI coding tools got better, everyone was encouraged to get comfortable with contributing in code. And we can thank our systems and operations team for setting up fantastic security and processes as well as documentation that's good enough that a non-engineer such as myself can spin up a working development environment. And last but not least, an incredibly helpful tool that I use every day for research and planning is Context AHC. It's an MCP server that gives AI tools access to all of the knowledge and data that we have documented over the years that I previously mentioned. So back to the experiment. About two months ago, we kicked off this project. The instructions were to pause the roadmap, work for 30 days, pair with a partner, and ship something real. We were welcome to use AI, though it wasn't a requirement. And of course, we couldn't all do this at one time. So we had to do this in stages and we had about a third of the company participating in the first round, that's around 500 people, that started around 794 projects in these 30 days. One lucky circumstance for me was that during this month, I was also participating in this two week in-person AI enablement course. So AI tools played a huge role in what I was able to deliver. And that's why I was building not one, but three different projects. And when Radical Speed Month started, I had some knowledge of GitHub. I would review PRs. I had played around with Cloud Code. But I had never actually built and shipped something in code at work. I've done a small hobby project here and there, but that was a completely different experience. I'm going to share my personal journey of how my skills and process and collaboration style shifted over the course of the 30 days. And this will also give us insights into how this change can be promoted and implemented at the organizational level for large companies. So this is the story of the first project I worked on and my first contributions in code. As I mentioned, at the time Radical Speed Month was starting, I was at the in-person AI enablement training. And over there, as part of the program, we did an exercise where as a group of four, we had two hours to build something together. We were two designers that aren't super comfortable in code, an engineer and a product lead. And we decided to build this board game session manager. We had a board game evening planned that night. So we wanted to create an app that's used to create, view and join and manage board game event sessions. You can see what that looked like. This is what we built in two hours. We built an app where people can start and stop a session, see what games and spots are available and even chat with each other. And on top of that, we chose this 16-bit style illustration of our actual office that people found especially appealing. The most interesting thing about this project isn't what we built. It was how we worked together and leveled up together. At the very beginning, we agreed that we all want to contribute and we were very lucky that the engineer that was in our group was really great at setting us up for success. He set up the project in GitHub, helped us grasp some basic commands and helped us understand how Git works. I worked with versioning in the past in different situations, and I thought I understood Git, but actually I didn't. I really needed someone to tell me how it works and what are the best ways to set up a project. We were sitting together, we were testing out the app and we were building it together and dividing the work among us. So we were all committing code to this project. And since this app that we were building wasn't related to any of our existing products, there were no risks involved and our focus was on learning, which was really great. We only used Cloud Code for this purpose. And outside of that, the only visual tool used was Nano Banana to create the office illustration. And the biggest insight for me from this project was that if you're an engineer, the impact that you have when you enable others may be far greater than the impact of doing more engineering yourself. And for any role, this is a great learning. It can be the same for engineering or design or product. Keep in mind that when you're working on a team with mixed abilities and if you're stronger in a specific skill, make sure that you're helping and enabling each other. Because now with AI, we can all do a little bit of the work that's outside of our domain. And this was the biggest learning for me, that engineers will need to become enablers and teachers. Moving on to the second project I worked on. This was my main focus for the Radical Speed Month. I paired with an engineer and a design operations person. And our main goal, we worked on some design system projects together in the past. And here, our main goal was to surface relevant and up to date design system information to humans and AI tools. And this problem space is huge. Our design system that we use for our WordPress ecosystem products is open source. It serves many products internally and externally and is always in flux. So the problem space was big. We saw a lot of opportunities. So we decided to divide and each person took on a problem to tackle. And what I proposed, this design system status tracker, the solution that I proposed raised questions with the engineer around performance, around maintenance, and if we would be able to build it in the first place. But I was empowered by the previous exercise. I was allowed to experiment. So I decided to build the full idea as a proof of concept on my own. So the discovery portion was three people, but the design and build was just myself. And it took around two and a half weeks to build this. And as I worked on it piece by piece, especially when I reached the point where I had actual live component previews working, I was really blown away at what I was able to achieve. I moved through the project without any restrictions in mind. I was just trying to push the limits. And the tracker was picking up links from the GitHub repo, from Storybook, from Figma, sorting them according to status, tagging them with the correct library names. Search worked perfectly. And I was really excited about this product. And I shared it with fellow designers. I iterated a little bit based on the feedback, but everyone was pretty excited. I built the prototype initially as a start. And then I used Figma only later in the process for some visual fine tuning, some things that I just couldn't explain to Cloud Code. I wanted to deliver an image and that worked. But the live project was what I was iterating on directly. And the prototype was built within a week. And then it took another week and a half to rebuild it for some fixes to set up hosting and deploy it in our internal platform as an internal tool, as well as on an external public facing site that had a little curated data and information. And delivering this product was a defining moment for me where I moved from a designer to a design engineer. This is, of course, enabled by AI, but most of all, it was enabled by the fact that I could own the entire process. And then, of course, within Radical Speed Month, I was allowed to experiment. I had the agency and made all of the necessary decisions. In large organizations, this is very hard to do. But here, I could just build whatever I thought was needed and I didn't have to spend a lot of time in negotiations and handover. And that was two and a half weeks. So I had one week left in Radical Speed Month. So I paired with a fellow designer. And here, you'll see the incredible shift in speed and process that happened. In only six days, we went from zero through the entire process to building an iOS chat for WooCommerce merchants that allows them to answer shoppers in real time, reply from their phone, or even set up AI to answer the questions for them. And this is a fully working proof of concept that includes authentication through WordPress.com, includes a Jetpack connection. If you use WordPress, you might know what that is. We also built a widget that inherits the site theming, where the site visitors can ask questions, as well as a fully capable AI agent that scans the site for information and answers visitors' questions in real time, and also discerns whether it can answer a question or not. What was important here was that the alignment and the ideation part was fairly easy. It was fairly easy for us two designers to align and shape the features because we always have the users in mind. And on the process side, what was interesting is that we started the exploration as a Cloud Code project folder, where all of the chats and ideas were recorded into a file system. And I noticed that significantly sped up collaboration, as well as the build later. And I've continued this practice in my daily work. The main artifact we worked on here as well was the prototype. And that's the biggest shift for our process because normally we would start in Figma and we would work in Figma until we reach very high fidelity. But here we had a very good plan. Then we built the prototype or the product. And then we went back into Figma to come back to it in a visual sense to build a mood board and for some UI fine tuning. And in only 30 days, my process that has been similar for years has completely shifted. And that was true for many other people. The first project showed me how the engineer's role can be elevated from a builder to an enabler. And this is a skill that I believe will be crucial as more and more people in other roles start working in code. The second one moved me from a designer to a design engineer. And here I'm talking about being able to push code to production in a well-established large system that requires serious onboarding and security practices. All of the company initiatives that I described together gave me the space I needed to experiment and find my own process that works within what we already have established in the company. And the third project that I shared was proof that we've unlocked speed because I already had adjusted my process a little bit, but then we both leveled up together and had a shared understanding on how we can apply the tools so we could also easily meet each other. If you're in a larger organization, try to implement this even just at your team level. Changing processes or tools in large organizations requires shifting the human behavior behind them. So provide your people with access to the new tools, find your enablers and champions, create space for experimentation and give them agency so that they can break out of their habits and unlock the speed of the tools. And then we'll see what AI tools can bring. Thank you.

0:08

SPEAKER_00

with it at our company that's somewhat unusual and very exciting. It was a one-month initiative called Radical Speed Month where two-person teams were given full autonomy to build and ship projects. I built three projects during this time. I will walk you through them and we'll talk about the outcomes and learnings for myself as a designer as well as the system-wide impact this experiment had. I'll tell you a little bit about Automatic first. We're the company behind several products in the WordPress ecosystem such as WordPress.com, Jetpack, WooCommerce as well as many other brands like Tumblr, Beeper and many more. We are around 1400 people and we're fully

0:56

SPEAKER_00

distributed which is pretty interesting for also what I'm about to show you. We're fully distributed and asynchronous which means that most of our work, our thoughts, our concepts from the past couple of decades are very well documented. I'm a product designer and I've spent over a decade building software and watching how teams build software. I've worked for, I've been at Automatic for five years now on the Jetpack design team and in the past I've worked for several companies of different sizes, different types of products and different team organizations configurations. So after all these years I've not only accumulated

1:41

SPEAKER_00

experience in designing products, I also have a lot of insight into team dynamics and organizational structures within product teams. Over the years the designer role has morphed and evolved a lot. We still have all of these different types of roles like web designers or UX designers or interaction or product designers. We've also used many different tools from the Adobe Suite to Sketch to Figma and many many others in addition to those. And AI is a new tool. It is definitely not built just for designers. It's versatile so it can be used by many other many different roles but the expectation is is the same for everyone. It's that it will bring us speed and of

2:34

SPEAKER_00

course this is why companies are rushing to put it to use. But no matter what role uses AI, this velocity is going to look very different in small teams like one to three people that are building new products, let's say a zero to one product versus large organizations that might have built entire product ecosystems that have software that's anywhere between five and twenty years old and that have more than a thousand or even thousands of people. Before I dive into the experiment, I will tell you a little bit about what preceded it in terms of how we use AI at Automatic. We were encouraged to use

3:16

SPEAKER_00

it, to learn it, and to experiment with AI from the very beginning. At first we had access to tools and courses. Then the AI enablement project was initiated where every single employee goes through a two-week course designed specifically for their role. This is completely immersive with lectures and focused hands-on time. Very, very helpful. And over time as AI coding tools got better, everyone was encouraged to get comfortable with contributing in code. And we can thank our systems and operations team for setting up fantastic security and processes as well as documentation

3:53

SPEAKER_00

that's good enough that a non-engineer such as myself can spin up a working development environment. And last but not least, an incredibly helpful tool that I use every day for research and planning is Context AHC. It's an MCP server that gives AI tools access to all of the knowledge and data that we have documented over the years that I previously mentioned. So back to the experiment. About two months ago, we kicked off this project. The instructions were to pause the roadmap, work for 30 days, pair with a partner, and ship something real. We were welcome to use AI, though it wasn't a requirement. And of course, we couldn't all do this at one time.

4:40

SPEAKER_00

So once we had to do this in stages and we had about a third of the company participating in the first round, that's around 500 people, that started around 794 projects in these 30 days. One lucky circumstance for me was that during this month, I was also participating in this two week in-person AI enablement course. So AI tools played a huge role in what I was able to deliver. And that's why I was building not one, but three different projects. And when Radical Speed Month started, I had some knowledge of GitHub. I would review PRs. I had played around with cloud code.

5:22

SPEAKER_00

But I have never until then actually built and shipped something in code. And I mean, at work. Like, I've done a small hobby project here and there, but that's completely, that was a completely different experience. I'm going to share my personal journey of how my skills and process and collaboration style shifted over the course of the 30 days. And this will also give us insights into how this change can be promoted and implemented at the organizational level for large companies.

5:58

SPEAKER_00

So this is the story of the first project I worked on and my first contributions in code. As I mentioned at the time Radical Speed Month was starting, I was at the in-person AI enablement training. And over there, we did an, as part of the program, we did an exercise where as a group of four, we had two hours to build something together. We were two designers that aren't super comfortable in code, an engineer and a product lead. And we decided to build this board game session manager. We had a board game evening planned that night. So we wanted to create an app that's used to create, view and join and manage board game event sessions.

6:45

SPEAKER_00

You can see what that looked like. This is what we built in two hours. We built an app where people can start and stop a session, see what games and spots are available and even chat with each other. And on top of that, we chose this 16-bit style illustration of our actual office that people found especially appealing. The most interesting thing about this project isn't what we built. It was how we worked together and leveled up together. At the very beginning, we agreed that we all want to contribute and we were very lucky that the engineer that was in our group was really, really great at setting us up for success.

7:36

SPEAKER_00

He set up the project in GitHub, helped us grasp some basic commands and helped us understand how Git works. I worked with versioning in the past for, you know, in different situations, but, and I thought I understood Git, but actually I didn't. I really needed someone to tell me how it works and what are the best ways to set up a project. We were sitting together, we were testing out the app and we were building it together and dividing the work among us. So we were all committing code to this project. And since this wasn't this, this app that we were building, it's not related to any of our existing products.

8:22

SPEAKER_00

So there were no risks involved and our focus was on learning, which was really, really, really great. We only used Cloud Code for this purpose. And outside of that, I think the only visual tool used was Nano Banana to create the office illustration.

8:46

SPEAKER_00

And the biggest insight I think for me from this project was that if you're an engineer, the impact that you have when you enable others may be far greater than the impact of doing more engineering yourself. And for any, this is a great learning for any role. It can be the same for engineering or design or product. Keep in mind that when you're working on a team with mixed abilities and if you're stronger in a specific skill, make sure that you're helping and enabling each other up. Because now with AI, we can all do a little bit of the work that's outside of our domain. And this was the biggest learning for me, that engineers will need to become enablers and teachers.

9:39

SPEAKER_00

Moving on to the second project I worked on. This was my main focus for the Radical Speed Month. I paired with an engineer and a design operations person. And our goal, our main goal, we worked on some design system projects together in the past. And here, our main goal was to surface relevant and up to date design system information to humans and AI tools. And this problem space is huge. Our design system that we use for our WordPress ecosystem products is open source. It serves many products internally and externally and is always in flux. So the problem space was big. We saw a lot of opportunities. So we decided to divide and each person took on a problem to tackle.

10:25

SPEAKER_00

And what I proposed, this design system status tracker, the solution that I proposed raised questions with the engineer around performance, around maintenance, and if we would be able to build it in the first place. But I was kind of empowered by the previous exercise. I was allowed to experiment. So I decided to build the full idea as a proof of concept on my own. So the discovery portion was three people, but the design and build was just myself. And it took around two and a half weeks to build this.

11:03

SPEAKER_00

And as I worked on it piece by piece, especially when I reached the point where I had actual live component previews working, I was really mind blown at what I was able to achieve. I moved through the project without any restrictions in mind. I was just trying to push the limits. And the tracker was picking up links from the GitHub repo, from Storybook, from Figma, sorting them according to status, tagging them with the correct library names. Search worked perfectly. And I was really excited about this product. And I shared it with fellow designers. I iterated a little bit based on the feedback, but everyone was pretty excited. I built the prototype initially as a start.

11:53

SPEAKER_00

And then I used Figma only later in the process for some visual fine tuning, some things that I just couldn't explain to claw to claw code. I wanted to deliver an image and that worked. But the live project was what I was iterating on directly. And the prototype was built within a week. And then it took another week and a half to rebuild it for some fixes to set up a hosting and deploy it in our internal platform as an internal tool, as well as on an external public facing site that had a little bit of a curated data and information. And delivering this product was a defining moment for me where I moved from a designer to a design engineer.

12:44

SPEAKER_00

This is, of course, enabled by AI, but most of all, it was enabled by the fact that I could own the entire process. And then, of course, we were within the Radical Speed Month, I was allowed to experiment. I had the agency and made all of the necessary decisions. In large organizations, this is very hard to do. But here, I could just build whatever I thought was needed and I didn't have to spend a lot of time in negotiations and handover.

13:22

SPEAKER_00

And then that was two and a half weeks. So I kind of had one week left, a Radical Speed Month. So I paired with a fellow designer. And here, you'll be able to see the incredible shift in speed and process that happened. In only six days, we went from zero to through the entire process to building an iOS chat for WooCommerce merchants that allows them to answer shoppers in real time, reply from their phone, or even set up AI to answer the questions for them. And this is a fully working proof of concept that includes authentication through WordPress.com, includes a Jetpack connection. If you use WordPress, you might know what that is.

14:13

SPEAKER_00

We also built a widget that inherits the site theming, where the site visitors can actually ask the questions, as well as the fully capable AI agent that scans the site for information and answers visitors' questions in real time, and also discerns whether it can answer a question or not.

14:38

SPEAKER_00

What was important here was that the alignment and the ideation part was fairly easy. It was fairly easy for us two designers to align and shape the features because we always have the users in mind. And on the process side, what was interesting is that we started the exploration as a Cloud Code project folder, where all of the chats and ideas were recorded into a file system. And I noticed that that significantly sped up collaboration, as well as build later. And I've continued this practice in my daily work. The main artifact we worked on here as well was the prototype.

15:25

SPEAKER_00

And that's the biggest shift for our process because normally we would start in Figma and we would work in Figma until we reach very high fidelity. But here we had a very good plan. Then we built the prototype or the product. And then we went back into Figma to just come back to it in a visual sense to build a mood board and for some UI fine tuning. And in only 30 days, my process that has been similar for years has completely shifted. And that was true for many other people. The first project showed me how the engineer's role can be elevated from a builder to an enabler.

16:16

SPEAKER_00

And this is a skill that I believe will be crucial as more and more people in other roles start working in code. The second one moved me from a designer to a design engineer. And here I'm talking about being able to push code to production in a well-established large system that requires serious onboarding and security practices. All of the company initiatives that I described together gave me the space I needed to experiment and find my own process that works within what we already have established in the company.

16:52

SPEAKER_00

And the third project that I shared was just proof that we've unlocked speed because I already had adjusted my process a little bit, but then we both leveled up together and had a shared understanding on how we can apply the tools so we could also easily meet each other. If you're in a larger organization, try to implement this even just at your team level. Changing processes or tools in large organizations requires shifting the human behavior behind them.

17:29

SPEAKER_00

So provide your people with the access to the new tools, find your enablers and champions, create space for experimentation and give them agency so that they can break out of their habits and unlock the speed of the tools. And then we'll see that AI tools can bring. Thank you. Thank you.

Reading tools

Type to find a passage

Appearance
Ask this transcript

Add a note