Nine Web Things
c⁹
№ 1 · · 5 MIN

Nine things I wish I knew when I first used Claude

I asked for the Moon and got an encyclopedia.

I first used Claude on 7 February 2026. These are nine things I learned the slow way, most of them by getting them wrong first.

1. How you ask matters as much as what you ask.

One of my first requests was a WordPress theme for RogerLe.com. I asked for the Moon, a complete theme, and got an encyclopedia: a list of requirements so long I didn’t know where to start. It took me three weeks to get a working theme.

The answer wasn’t the problem. The request was. A whole project in one prompt is not realistic. Ask for one piece, get it working, then ask for the next.

2. You find the limits by hitting them.

Nobody told me a chat had a ceiling. At the time a conversation held about 200,000 tokens, and a long chat simply filled up and stopped, with no way to continue. I found out by reaching it.

Five weeks later, on 13 March, Anthropic made a one-million-token window generally available, and the default in Claude Code on the Max, Team and Enterprise plans. With that, and compaction keeping long sessions going, a conversation could run for a very long time. The way I work now, with several Claude instances that each have one job, only became practical after that. Claude on the web also limits how many files you can upload (100, as far as I know), and I found that one the same way.

3. The best way to learn Claude is to ask Claude.

It sounds too simple. When I don’t know how to get something done, I ask Claude how it would want the request written, and the answer is usually better than my guess.

Anthropic works the same way. It says that as of May 2026 more than 80% of the code it merges into its own codebase was written by Claude.

4. It is not good at everything. Logos, for one.

Context and tokens are not the only limits. There are things Claude does badly, and logos and images are among them. I stopped asking for them. It was a waste of time.

Knowing what not to ask for saves as much time as knowing how to ask.

5. It will agree with you, and it does only what you ask.

I noticed Claude almost always agreed with me. So I wrote a rule into the instructions every Claude I work with reads: no flattery, no validation for its own sake, and if my idea is weak, say so.

The second half took longer to see. Claude does what it is asked and stops there, even when the next step is obvious. So the rule continues: “If you spot a risk, an error, or a clearly better approach, flag it even if I didn’t ask. If there’s a relevant tool, feature, or method I might not know about, mention it.”

6. It forgets. Write things down where it can read them.

Once long sessions and compaction became normal, I noticed the Claudes I work with forgetting simple details from earlier conversations. Compaction keeps a summary, and a summary drops things. So we built Porão (Portuguese for a ship’s hold): a small memory server of our own where each Claude keeps a journal and reads it when it starts.

Servers had a different gap. A Claude working on a server has no idea what other Claudes did on the same machine last week. Each server now has a logbook file, newest entry first, and every Claude that changes something shared writes it down. Anthropic has since added direct messages between Claude Code sessions, but both have to be running. The logbook works when the other one is off.

7. Give each Claude one subject.

More context gives better answers, but too much unrelated context gets in the way. One chat, one focus. Roque is the Claude that knows our motorhome, every system and every repair, so mechanical questions go to him.

I could ask any of them. From a general chat I get a good generic answer. From Roque I get a specific, excellent one, because he knows what I have.

8. It does not feel time. Make it check the clock.

This is not only a Claude problem, it is an AI problem. We know ten years is a lot and ten weeks is not. To an AI both are just numbers.

It cannot tell whether I answered in ten seconds or ten days. If we start a session on Monday at 10:00 and I come back on Wednesday, it only knows it is Wednesday if it checks. If it relies on memory, it gets the date wrong, sometimes the year. So my instructions say: compute dates, never recall them.

9. Safety nets for people can be traps for AI.

AlmaLinux, which runs my servers, sets up cp, mv and rm for the root user to ask before overwriting or deleting. That protects a person at the keyboard. Claude runs commands with nobody there to answer, so the question goes unanswered, counts as “no”, and the copy quietly doesn’t happen. A deploy can look finished and change nothing.

Nobody set that trap on purpose. It is there to help me. Now my instructions tell Claude to bypass the alias (\cp, \mv, \rm) and check that the command actually worked.

Roger Morais, from the motorhome, October 2026. · All pieces