Finding My Spark Again: A Month With Codex
I thought this would be a post about Codex. It turned into a post about burnout, identity, and finally enjoying building again.
In this post
- The before
- The spark
- Building without pressure
- AmpCode and ClawdBot
- Enter Codex
- The planning phase changed everything
- From builder to conductor
- The agent swarm
- The Codex app unlock
- What changed in me
- Links and references
The before
I have been burned out for years, longer than I wanted to admit. I worked at Zendesk in 2015 with some of the smartest, most passionate people I had met, and I did not handle that period well. I was young, loud, and bad at navigating US leadership dynamics. I handled conflict poorly and got let go. Since then I have had many jobs, but none felt like that team. Ten years later, I can see I was still chasing that same feeling.
I also learned I likely have autistic traits, and that explains a lot about how corporate ceremony hits me. SAFe-style process stacks, meetings about meetings, follow-up rituals for the ritual itself. It drains me fast. During those years I kept working festival medical teams in Denmark, and that helped me stay afloat. Music helped too. Most days I run 8-10 hours; lately it has been heavy melodic early-2010s hardstyle.
A few years ago I moved into festival operations and saw teams still running incident work from one shared Excel sheet. It technically worked, but barely. I hated it, so I built Shamira, a Rails system for incident logging and radio coordination. ActionCable made the real-time part easy enough to ship. I built that while consulting for large Danish companies in DevOps/SRE/platform engineering, whatever term was fashionable that quarter.
Shamira started hand-built, then summer 2025 became a tooling experiment. I tried Cursor and copied the PRD-heavy workflow from Ryan Carson’s How I AI interview. It worked, but felt clunky. A month later I switched to Claude Code, started small, burned through tokens, upgraded, and kept pushing. The PRD style still did not fit my brain; I ended up with 50+ markdown files Claude maybe read and maybe ignored. Code quality was okay, and Rails testing culture saved me because strong system tests let me refactor without pure fear.
My GitHub graph for June and July looks productive, but it was unsustainable. Long days, blurred focus, Path of Exile on one screen while Claude churned on another. When festival season ended, I crashed. My girlfriend and I took two fully offline weeks at my parents’ vacation house in a quiet part of Denmark: nature, books, jigsaw puzzles, no feeds.
I also returned to CrossFit. I started in 2013, lost the routine after Zendesk, and in August I finally found a small gym with the same family vibe as the old days. I have logged almost 90 sessions since then, and it is the best burnout intervention I have tried.
The 2025 market was rough for Danish consultants, so we lived off savings for a while and planned for a long runway. Then I landed a contract in October and pressure dropped enough to breathe again. Still, fall 2025 left a mark. I hated building, avoided side projects, and every thought about Shamira came with guilt. I pushed everything aside, avoided customer calls, and admitted publicly that building felt scary.
The spark
A few days after New Year’s Eve, I spoke with my childhood friend @madsmadsdk, who was shipping, posting, and actually having fun. I asked how he kept momentum without turning everything into a chore. His answer was simple: write quality replies on X every day. I started smaller, just 10+ good replies/day, and that tiny habit gave me my spark back.
Funny how opening X at just the right moment can pull you out of a year-long burnout spiral
I was suddenly back on X daily, reading about people building things. Not pressure to ship, just curiosity to explore. For the first time in years, I wanted to try something new.
Building without pressure
I kept the first step small: personal site + library. I wanted book pages to be raw notes, not polished essays. Read the books, do the exercises, log what I learned. I started with K&R C and focused on re-learning how to learn.
Worked through K&R’s C book over a weekend
It helped a lot. Memory and algorithms felt fun again, to the point where I tried explaining one to my girlfriend. She hated every second of it. Around the same time, training started clicking again too.
First strict pullup from dead hang
At the same time, everyone in the X tech bubble talked about AI coding agents: Ralph, AmpCode, Codex, Claude Code, OpenCode, and the whole Claw ecosystem. I was skeptical. “AI will replace developers” takes were everywhere, and in June and July I still had to heavily steer everything. Still, I was curious.
Finally tried the Ralph/AmpCode thing
First impression was heavy, token-hungry, and awkward, but also different in a way I could not ignore.
AmpCode and ClawdBot
After a few days, AmpCode clicked. I had wanted to move toward a leaner 37signals-style approach, but I was nervous because Shamira had a big RSpec + system test setup. Then I did a longer planning session and let it run:
AmpCode converted my entire test suite
The refactor pace improved immediately.
And the refactor was much faster
For the first time since August, I felt momentum instead of dread.
One late night I hit an AmpCode easter egg where the model said “You’re absolutely right.” I laughed so hard I woke my girlfriend up.
That moment mattered because I was having fun again. I have tried every productivity system under the sun and none stuck for long, but then I tried ClawdBot on Telegram, initially just for journaling. It became one of the most useful burnout tools I have used: commits from GitHub, activity from X, training logs, check-ins during the day, and a shutdown routine at night that showed what I had actually done.
I had been lying to myself for years about being “unproductive.” I kept ending days with this vague sense that I had done nothing, even when I had spent hours shipping, training, or helping people. The logs showed the opposite.
Around the same time, I started reading how people who were really getting leverage from Codex worked in practice. It was not just “write a better prompt.” It was a proper AGENTS.md, clear conventions, and tool wiring that removed friction and made good output repeatable.
That sent me down the rabbit hole
That also forced me to clean up months of Claude-generated clutter.
Then I got my first real one-shot:
It understood the assignment and built the whole thing. I was not “fixed,” but I was building again.
Enter Codex
People kept saying Codex was different: slower, better quality. When I finally tried it, that was exactly my experience. It was slower than AmpCode, but output quality was more consistent and needed fewer revision loops.
I read @kr0der, adjusted prompts, and gave Codex more time to plan.
Right after that, I added the ask-questions-if-underspecified skill, and it changed my day-to-day flow immediately. It forced clarification before implementation when requirements were fuzzy, and it cut a lot of rework and wrong turns.
Codex patterns I noticed were different from how I usually structure code; not always my style, still effective. I let Codex run while I read. That was new for me. I was delegating technical execution instead of micromanaging every line.
The planning phase changed everything
I asked Codex to integrate Whisper speech-to-text into Shamira. I expected a long slog: audio quality issues, noisy environments, Rails integration details. Instead, one-shot. Whisper-large-v3, background noise handling, clean Rails integration.
That day changed how I prompt. I now front-load context, spell out constraints, map the architecture, and include concrete examples before asking it to write code.
Give Codex 10 good minutes up front and it can save you hours later. AmpCode tolerates vague prompts. Codex punishes them, and that pressure made me think more clearly before I build.
From builder to conductor
I had two Codex instances running on different parts of Shamira, one feature track and one refactor track, and that was the point where the hard part stopped being coding and became orchestration. I had to figure out how to avoid conflicts, keep context straight, and still move fast without creating cleanup debt. The setup itself stayed simple: multiple clones, one agent per folder, clear ownership, reset flow after PR merge.
Then I found a self-improving AGENTS.md pattern, and things became much more coherent. Prompts got longer, but output quality and consistency improved a lot. I also explored marketing skills from skills.sh, then had Codex extract what was relevant for Shamira. That helped me work through areas I had avoided, especially SEO, positioning, and messaging. Codex did not “do marketing for me”; it gave me a structure to think with.
That same day I had a real customer meeting. Stressful, but good. They wanted what I was building. Right after that came Rails extraction, where I finally had enough patterns to codify what mattered most: convention over configuration, sensible defaults, and structure that guides without handcuffs. That mirrors what I have always liked about Rails, and it gave me a reusable system tuned for my brain and my workflow.
The agent swarm
Around the same time I expanded OpenClaw into specialized agents. OrcDev released AgentSkills, I installed it, and started composing a smaller team with clear responsibilities instead of a giant bot zoo.
The core setup now is Claire as executive assistant, Leo for research, Marcus and Elena for technical leadership and review, Noah plus Nina and Mason for training/nutrition/sleep coaching, and Sophia for finance. Telegram limits forced consolidation, which turned out to be good for quality. Fewer bots, clearer ownership.
The daily logs are long, but useful. I get better sleep habits, better nutrition support for training volume, a clearer view of burn rate and runway, and steady business learning.
The Codex app unlock
Later, OpenAI released the Codex app, and it made the multi-run workflow much easier for me because I no longer had to juggle as much manual setup for parallel sessions.
Under the hood it uses Git workspaces in a really smart way, so I can run multiple Codex tasks at the same time without stepping on my own changes. It feels smoother, more focused, and much closer to how I naturally work now.
What changed in me
Codex was not the full story. The real shift was identity:
- from “I have to do everything myself”
- to “I can design systems that help me do better work”
From skeptic to conductor is less about AI hype and more about control. I choose the direction, set guardrails, review the output, and keep the responsibility. Honestly, it helped me trust myself again.
My practical takeaway from this month is simple: start smaller than feels impressive. One page, one prompt, one focused hour, then repeat. That is what helped me build momentum again.