When I started building trustedone.pl with Claude and Codex, I thought it would be fast. I type a command, AI codes, I fix the details. Reality turned out more complex – and far more interesting.
The paradox of powerful tools
The better the AI, the more you need to supervise it. Sounds absurd? After two months of intensive work on the site, I can say this is the most important lesson of the entire project.
Claude has a context window of ~200 thousand tokens. A lot. But a debugging session can burn through those 200k in three hours – and then the model starts “forgetting.” It repeats the same mistakes. Tries solutions that already failed. Goes in circles.
The solution? Documentation. Not as bureaucracy. As long-term memory that AI inherently lacks.
The rule that changed everything
70% planning and documentation, 30% actual coding. Sounds like wasted time? I tested this on real tasks.
Adding RSS feeds to geographic sections on the news page. Without documentation: 4 hours of frustration, partial solution, three sessions debugging the same problem. With documentation: 2.5 hours, solid solution, zero returns to the issue.
The difference? I created a task specification before coding. Problem description, data structure analysis, specific lines in files to modify, success criteria. 148 lines of documentation for 3 code fixes. Codex completed the implementation in 45 minutes. Zero errors. Nothing broke in the process.
AI drift – when the machine starts wandering
There’s a phenomenon I call “AI drift.” You ask the model to fix the page layout. It starts with the template. Soon it’s refactoring CSS. Then JavaScript. Eventually it proposes rebuilding the blog template. Two hours gone. Nothing is fixed.
A plan prevents drift. Not an elaborate system – a simple document with a list of steps and estimated time. Diagnostics: 10 minutes. Icon fix: 20 minutes. Staging test: 15 minutes. Total: 75 minutes. The model has guidelines. I have control.
When to close the session and start fresh
Continuing a long session seems logical – after all, AI “knows the context.” Practice shows something different. Efficiency drops exponentially. After one hour, decisions are good with 10 thousand tokens of context. After three hours – poor decisions with 180 thousand tokens of load.
Symptoms of overload? The model repeats the same solutions. Proposals become overly complicated. Simple problems grow into projects.
Strategy: after three hours – stop. I document what I tried, what worked, what didn’t. Close the session. Open a new one. A fresh Claude instance reads the notes from the previous session and solves the problem in 15 minutes. No baggage.
Numbers from two months
The trustedone.pl project in statistics. 12 thousand lines of documentation. 50 hours to create and maintain it. 93 hours saved thanks to it – prevented deployment errors, faster task handoffs between sessions, avoided debugging loops.
Balance: 43 hours gained. 46% return on investment in documentation. Deployment errors: from 5 to 0. Percentage of work requiring rework: from 25% to 5%.
What this means for communication
Years in PR and communication teach you one thing: brief before production. Strategy before tactics. Goal before tool. Working with AI follows exactly the same logic in a new context.
A language model won’t replace thinking. It speeds up execution – but only when you know what you want. Documentation is the brief. Plan is the strategy. Code is the tactics.
Vibe coding sounds like improvisation. Good vibe coding is prepared improvisation – like jazz, not chaos.
I’m building trustedone.pl as a place for reliable information and global inspiration. If you’re interested in the practical side of working with AI in content creation and projects – get in touch.