Real Lessons from Claude Code, Codex, and AI Agent Tools

This is a summary of nearly two months of intensive use of Claude Code and Codex. I'm pretty sure something in here will be useful to you, whatever your background.

Even if you have zero technical training and have never touched AI beyond a chat interface, I think this will give you something new. I was exactly that person not long ago.

One thing to be upfront about: Claude Code has done a lot of things for me, but it didn't write this article. Though it did offer edits and suggestions along the way.

How It Started

It all started because I genuinely couldn't stand the mindless searching and boring organizing anymore.

I got interested in startups and started cold-emailing CEOs in March to pitch myself. The process was exactly as tedious as it sounds: go to a VC's website, click through their portfolio one by one, search the CEO's email, write the message, copy everything into a Notion database, paste it into my email client, send. The constant context switching was exhausting.

One day I just couldn't take it anymore. I asked chatgpt for help automating some of it. It suggested n8n. I got a Reddit scraper working, but the full cold-email workflow refused to cooperate no matter what I tried. I was venting to a friend when he said, almost as an aside, "Codex and Claude Code are really powerful now." I tried it out of pure mind(just give it a try anyway): I described my exact workflow to Codex in detail. Six minutes later, it had built something that worked better than what I'd been doing manually.

The feeling was hard to describe. I imagine it's something like what a textile worker felt seeing a spinning jenny for the first time: shock, a little fear, and then one immediate thought: "I really need to understand this." It felt like a watershed moment.

I started hunting for tutorials. Claude Code had far more documentation than Codex since it launched earlier, so that became my entry point into what turned into months of intense exploration.

Exploration

I started throwing everything tedious at these tools to see what stuck.

My meeting notes with my supervisor were never complete enough. I wanted a specific format: direct quotes, highlighted key points, clear next steps. So I had Claude Code lock in an SOP format. Now every meeting produces a perfectly structured note automatically.

Presentation slides were always a headache for me. Using good-looking HTML templates and content I provided, I worked with Claude Code step by step to turn ideas into HTML and then PDF. Claude Code wrote all the code. After many rounds of back and forth, I also locked in an HTML Slide SOP. Now slides can generate automatically.

I had ideas I wanted to track in a database, but opening Notion and clicking through to the right database always took too many steps. So I set up a Telegram bot hosted on Railway: now any idea I have lands in the right Notion database with a single tap. Claude Code wrote all the code.

Before, when I found good content online (say, a YouTube video), I'd copy the transcript, hand it to an AI for a summary, specify the format every time, and paste the result into Obsidian. Now the Obsidian Web Clipper plugin pulls it in automatically. An LLM formats it as a wiki-style note. When I want to go deeper on something, I ask the AI on the spot and update the note based on the answer. Then I use the Feynman method (which I've locked in as a reusable skill) to test whether I actually understood it.

There are so many more examples. The pattern is always the same: have an idea, act on it fast, learn from what breaks.

Lessons Learned

When you experiment enough, you see these tools clearly. They're powerful, but they have real flaws. You won't always get what you want. You have to manage them.

It Only Cares About Right Now

AI solutions have one obvious default: make it work now, worry about nothing else. Long term maintenance isn't on its agenda. It prioritizes the immediate problem over architectural durability every time.

My fix: after getting any proposal for a long-lived project, I now add one follow up: "Is this maintainable long-term? Is there a simpler way to do this?"

Backups Come First

I once read a Reddit thread where someone asked Claude to help with a task, and Claude deleted their entire database. Just gone. You can't watch every action it takes, but you also can't afford to babysit it constantly.

The answer is Git. My Obsidian vault syncs to a private GitHub repo via the official Obsidian Git plugin. Code projects follow the same pattern. If you don't have a development background, you don't need to learn much. One afternoon to understand commit and push is enough. Claude Code wrote my auto-backup scripts.

Never Hand Over Credentials

One thing that genuinely frustrated me: it'll ask for your API key directly in the chat.

Large models aren't well-trained on security, so you have to protect yourself. API tokens, passwords, credentials: none of these belong in a conversation window. Create a .env file yourself and keep sensitive information there.

CLAUDE.md Needs Regular Pruning

The more you use the tool, the more rules and preferences you pile into CLAUDE.md. But once it gets long enough, Claude stops following all of it. There's an effective ceiling on how many instructions it actually processes, and rules you added quickly sometimes contradict the overall logic.

Review it regularly: delete what's outdated, merge what's duplicated, keep it lean. For any configuration file, less is more.

Treat It Like a Junior You Have to Manage

This is the most accurate framing I've found. You have to actively manage it, not just dump a requirement and wait.

My current workflow:

Set a clear goal first. Not "help me do X," but "the goal is Y, it needs to reach Z, and here's what done looks like." A fuzzy goal makes everything downstream useless.

Make it research before building. Most things you want already exist on GitHub or in some community. Search first, reuse if you can, don't start from scratch.

Require a plan before execution. If something genuinely needs to be built, ask for a plan before any code gets written. This looks slow and actually saves time.

Keep progress visible. I have Claude maintain a LOG.md. Each completed step gets a checkmark; errors get logged. I don't need to watch constantly. A quick glance tells me where things stand and whether it's stuck.

Dig into the root cause when it loops. If the same problem comes back three times, stop. Don't ask for another proposed fix. Make it understand the actual problem first. Throwing more solutions at a bug it doesn't understand will not unstick it.

Require verification, not just "done." "Done" doesn't mean done. I now explicitly ask for tests and verification. The result counts only when I see it actually run.

Lock working patterns into SOPs. Every time I solve a new type of problem, I run a quick retrospective with Claude and write the outcome into an SOP in Obsidian. Next time a similar situation comes up, I hand it the SOP. No re-explaining from scratch.

Resource Recommendations

If you don't want to hunt these down yourself, I've organized everything into a categorized database. Claude Code organized it, naturally.

Tool Resource Library (Categorized)

A few worth calling out:

Superpowers Skills: a skill pack for Claude Code with built-in workflows for brainstorming, task decomposition, and execution planning. Having a structured approach beats having none.

Claude Mem: cross session memory. By default Claude starts fresh in every conversation. This changes that.

Context Mode: automatically manages context, extracts relevant information, saves it locally, reduces token usage, and keeps Claude better oriented to the current task.

Agent Style: an English writing ruleset that strips passive voice and AI-register patterns from Claude's default output.

Typeless and Wispr Flow are both my referral links. Not sponsored. Both give a free month to new users. Neither is a simple transcription tool. Both strip filler words, restructure content on the fly, and handle mixed-language input well.

Wispr Flow runs globally on Mac with a lighter footprint. Typeless does more active work: it takes scattered speech and organizes it into coherent, structured text. The tradeoff is price. I currently prefer Typeless, though Wispr Flow is a solid alternative. I've tried other tools too, and once you've used the better ones, the gap is immediately obvious. Voice input alone can meaningfully lift your output.

On Cost

I started on the Pro plan ($20/month). I found myself constantly checking the usage dashboard, watching the remaining quota, occasionally hitting the 5-hour cap. Eventually I upgraded to Max. I haven't looked at those numbers since.

That said, you don't need Max from the start. Codex's free tier is generous. The exploratory phase is totally manageable on free tiers. If Claude Code later becomes your primary tool, Max starts making sense on its own: your quota directly limits your output, and the ROI is high.

What I'd Leave You With

AI hasn't reduced my screen time. If anything I spend more time at my computer now.

What it changed is how learning feels. Before, it had real friction: no immediate feedback, no way to know if I'd actually understood something, constant page-switching and copy-pasting. Now I ask anything I don't understand on the spot, update my notes after understanding it, and test with the Feynman method. The whole chain connects without ever leaving the current page.

The other shift: ideas move from thought to action much faster. Before, getting from "I want to do this" to actually doing it meant crossing a lot of terrain. Now I just talk to Claude: can this be built, has someone already done it, what are the steps? Then I start. That loop is genuinely addictive.

One thing that used to stress me out: this space moves fast. For a while I felt perpetually behind, finishing one tutorial only to see the creator post an update. I've made peace with that. Anxiety helps nothing, and most of the latest features are still being iterated on anyway. What actually compounds over time is more fundamental: engineering thinking, management thinking, and making your workflows repeatable.

The principle I hold onto most is that workflows should be portable. Claude Code is excellent, but all your data and configuration lives on Claude's platform. If Claude stops being the dominant tool someday, you need to be able to move your whole workflow somewhere else. That's why I keep every preference, SOP, and writing style documented in my own Obsidian vault. When I eventually switch tools, I give the new AI these files to read. It learns who I am and what I want. The tools will change. What you've built for yourself stays yours.

This is the best of times. I still believe that. Keep exploring as always.