I love writing: code, documentation, tickets, copy for landing pages… I used to write dozens of pages per week as a teenager. There’s something about writing that’s liberating. Ideas, doubts, technical challenges. If I can sit down and pour it into a notepad, it does not solve the problem, but at least it takes it out of my head, which makes room for solutions to come and visit.

There are scientific studies suggesting a strong correlation between the habit of writing and improved critical thinking and problem solving skills, and I vouch for it, because writing is thinking. Not the “I need to patch and fix this UI production bug urgently” thinking, but a slower “let’s go on a hike and enjoy the view” thinking. It forces you into Daniel Kahneman’s System 2.

Recently, a few coworkers were struggling to make Claude Opus 4.8 generate compliant video scripts for an overdue marketing campaign. Management was pushing an aggressive deadline, and the guys were doing their best to try and produce something breathtaking, short, and fast. And we all know Claude can generate a good output fast, but fast does not guarantee greatness, “good” does not mean “right”, and I do not think it was the right approach at all to brainstorm (but that will be a separate conversation).

I refuse to dramatize the event, so I’ll cut the chase.

There were 3 foundamental problems:

• The deadline was perhaps a little too aggressive, not leaving room for proper ideation and creativity.

• The tooling was not ideal. Claude is great for generating high-quality output if the context and instructions are granular and leave no room for weird interpretations. I’ve had better results brainstorming and refining requirements with GPT 5 (for code-related tasks) and Gemini 3 (for creativity-related tasks).

• The prompt lacked all the context my marketing coworkers KNOW about the company, our language, voice, design system, how they do ads, keywords… Claude knew close to nothing, yet they were expecting an output fully aligned with our brand.

Given the circumstances, I could not do anything about #1 and #2, but I knew #3 could be drastically improved; we just needed to quickly fix a set of wrong assumptions.

Assuming that feeding a basic prompt like “Generate a JSON script for a 30-second video of {company_name} with a strong hook and smooth outro where the main topic is {campaign_topic_details}” will generate a decent video is plainly wrong and is full of holes.

Having worked closely with video editors for a few years, I know first hand there’s so much that goes into crafting a 30-second video, and probably it’s so second nature to the team that they just assumed “Claude will get it”. They were expecting that, with a simple prompt, Claude would crawl through our various websites, map all our marketing resources, explore the Meta Ad Library, and crawl through our Trustpilot reviews to find the right keywords and craft a compelling message and a brilliant video. All of it in just a few minutes. The problem is, LLMs don’t “take care of it”, they fill the gaps and produce an output. They’re lazy, by design (optimizing cost), so don’t expect any crawling or data gathering unless you explicitly ask for it and provide sources.

The solution? Fill the gaps yourself, so that you let little to no space for hallucinations and missaligned specs.

I recommended just that, but it wasn’t fun. Because it requires writing an extensive set of rules before the task even starts, and writing an extensive document after having had an AI writing in your stead for months just feels like a huge labor-intensive chore (that’s why so many skip it).

Of course, there are ways of providing all this context in a consistent basis and less overhead, like creating custom skils, markdown files and such, but it all just means you do the heavy writing once.

So, back to our story, what happened then? Nothing. This is not a success story telling how I “helped drive workflow efficiencies in creative output”; it’s a personal example that illustrates why I’m covinced writing is important.

As it happens, making mental space for writing seems to be a big challenge. They thought there was no time for it, so they thanked me and promised to come back for help in writing the context files and skills later (Spoiler, we did and later on, results were great).

But for this ocassion, they just kept reiterating over the output until something satisfactory enough came out. I’m happy for their hard work, and I wish we all could get into the habit of writing early into a task.