Senior Software Engineer, Frontend Core · Flock Safety
Flock Safety builds public-safety technology. The microfrontend shell is the web app that every product team's app runs inside.
As the only developer on the team, Cory spent the year building the things other teams ship on: a microfrontend shell, a notification platform, a service template, and a scheduling service. A founding member of the Frontend Guild and its lead in Q2, Cory fully launched the shell in mid-August, so every user of those apps now goes through it. The share of page views with an error, across every app in the microfrontend, fell from about half to under a third.
AI is the newest change, and the steepest. For most of the year it was an assistant in the editor (Cursor, with Codex on the side). It moved to Claude Code in June. Since late summer Cory has been building AI tooling for other people: a skill for teams that consume the notification platform, handoff agents that carry the shell's know-how, and an epic builder that turns notes into planned, Jira-synced work. Most of that is weeks old, and this report says so where it is.
To me, scaling yourself means two things: growing in the technology I work with, and raising the bar for everyone around me. I want the floor for what's acceptable to keep rising, and I want it to keep getting easier to reach. That's why I build tools, services, and packages that anyone can pick up and use. A rising tide lifts all boats.
What surprised me most this year was how much AI can do. A year ago I used autocomplete and the occasional chat window. Now AI is part of how I review code, write tickets, and generate whole features. I was reluctant to hand off the actual coding, but orchestrators, skills, and agent files changed my mind about how accurate the output can be. AI isn't perfect, and I still want to be able to defend every architectural choice. I just don't think every line of code needs to be fought over anymore.
If I could replay the year, I'd have started with orchestrators sooner and built more tools for my own workflow earlier. The developer dashboard I've been extending to write epics and work through tickets has been a big boon.
Every chart below shares the same 12 months. Markers show the three turns in the year: guild lead in April, the move to Claude Code in June, and general availability in August.
Each month written up like a release note. The stats line is the same every month: PRs merged, reviews given, tickets resolved, meeting hours, and AI prompts sent.
Cory worked in 17 other teams' projects, repos, or design docs this year, and did substantial work for about 9 of them. Two of those 9 are teams Cory also belongs to. The count is generous: a single review is enough to include a team, and the shell rollout into other teams' apps counts too, though that was Cory's own project.
In March Cory built a Node/React service template (11 of its 13 commits). In May a sibling team proposed a PDF-generation service, and Cory approved the design. A new Engineer II on that team generated the service from the template and built it, writing 46 of its 52 commits. Cory's part was review and guidance rather than code.
How: mostly live Slack huddles walking through the code, with review as the backup. The focus was patterns and maintainability: bringing the service in line with company standards so the next engineer could understand it right away. It also covered how to get good results from AI by giving it the right context.
Beyond mentoring, Cory is a founding member of the Frontend Guild (two years running) and led it in Q2, approved five other teams' design docs, and wrote runbooks and setup guides so teams could handle recurring requests themselves.
Mar 2026Template built. CI, deploys, and conventions baked in for any new service.May 2026Design approved for the sibling team's PDF service.Jul 202620 review passes as the first code landed.Aug 202640 review passes, 81 inline comments. v0.1.0 shipped Aug 25.Oct 202624 releases and 25 production deploys later, the mentee owns it.Real-user monitoring for the microfrontend shell, from the dashboard and alert Cory set up in July. The figures are relative, so no business numbers are shown. Monitoring started mid-July, so earlier work has no baseline.
Largest Contentful Paint (p75) stayed around 8 seconds all period, which Google rates poor. Load time (p75) rose from 3.3s to about 4.0s during GA as more apps moved onto the shell.
The dashboard already breaks load time down by remote app, by asset origin, and by network speed, so the work starts from data. The error drop above is for all apps combined. It coincides with the late GA fixes, the move to pnpm, and more apps joining the shell, and the data can't separate those.
Three tools and three ways of working: pairing in the editor, delegating in the terminal, then building agents and skills for other people to use. Most of the change happened in the last three months. AI hasn't clearly raised output yet: commits doubled in April, before Claude Code, and October's merged PRs were half of July's. What changed is the kind of work.
Cursor in agent mode as the daily driver, with Codex for heavier component-library work.
Claude Code from June 5. Cursor stayed as the editor; agent work moved out of it.
Opus for planning, Sonnet and Haiku for the work, subagents in parallel. The new part is building agent tooling for teammates.
The problem. Cory was mostly working solo, and the planning showed it. Epics were a single sentence, tickets were thin, and nobody was reading them. The plan lived in Cory's head, which made the work hard for anyone else to follow or to look back on later.
The build. In one day, Cory added an epic builder to a teammate's AI developer dashboard, working the way all of Cory's agent work goes now: design doc, then plan, then agents build it, then review. That came to 22 commits and about 4.9k lines, with tests. It writes to Jira through a three-way merge preview, so nothing changes without Cory seeing it first.
“I can think out loud into the tool, and it pieces it all together, does the research, and writes a cohesive epic and the tickets.”
The result. The same day, a one-line idea became a researched epic with 19 scoped tickets: the plan for handing off everything Cory owns. Managing the work got easier, and anyone can now see what was actually done under the epic.
Status. It's in Cory's local fork for now. The next step is a PR to the original tool so the whole team can use it.
mfe-mui-context-boundaryA skill that diagnoses a silent microfrontend bug: UI components that render but never update state, with no errors anywhere. Now being added to the shell repo for everyone.Jul 2026 · PR opennotification-center skillA skill merged into a consuming team's repo, so their agents know how to integrate the notification platform without asking.Aug 2026 · mergedhaiku-shuntA plugin that sent bulk file reads to a cheaper model. Measured it, saw total token use go up instead of down, and turned it off the next day.Sep 2026 · retiredepic builderTurns notes and images into a researched epic and planned tickets, synced to Jira. See the case study above.Oct 29 · local forkadd-remote · mfe-debugger · mfe-expert · mfe-platform-guideA skill for adding an app to the shell, plus three subagents that carry the shell's know-how, so others can work in it without its owner.Oct 28 · PR openhandoff auto-refreshRegenerates those subagents' knowledge after every merge, so it doesn't go stale.Nov 2 · PR openWork rhythm from AI prompt timestamps, the calendar, and Slack. Times are US Eastern.
I learn best from structured courses. I watch a module, then try it by hand before I look at the answer. If I get it right, it feels like my brain grew. If I don't, it isn't a defeat; I'm trying something new and getting better. Either way, I study the real answer until I understand why it's right.
Lately that's how I've been learning game development, on GameDev.tv. The C# combat library under Outside work came out of one of those courses.
Topics that came up again and again in AI sessions, design docs, and side projects.
Personal projects, mostly August to October. The common thread is trying an idea with agents to see if it holds up.
Claude "portfolio managers" with different strategies trade simulated portfolios against buy-and-hold and random baselines. A second round reads a briefing a cheaper model writes, to test cost against quality.
A browser version of a favorite bag-building board game, with lobbies and invites, for playing with friends.
A paper-trading bot that watches dated crypto futures against their index price and simulates cash-and-carry trades when the spread is wide enough.
A combat and damage-event library with a sandbox, plus a small game prototype built on it. Built while working through a GameDev.tv course.
A personal site with no build step and no dependencies, updated as the year goes.
One Claude Code session, directed by Cory. Eight background agents pulled the sources in parallel while the main session asked Cory questions, filled gaps, and wrote this page. The agents did the gathering. Cory made the calls on what was true, what was fair, and what was safe to share.