What You Asked For,
and What Actually Ran

Claude used MCP servers for me one thousand three hundred and five times over forty days. I never once asked for it, because I did not know it was there.

That is the whole problem, and it is easy to miss because nothing breaks. The output is good. You move on.

Then the models get better. So you ask for more: build the whole app, test it, deploy it. Now twenty-three things happen. You still understand three.

The gap does not close as the tools improve. It widens.

The number of steps you understand depends on you, and it barely moves. The number of steps executed depends on the model, and it climbs every release. Every capability jump makes the delegation bigger and your share of it smaller.

The evidence

My own forty days, read backwards

I scanned my Claude Code history: 26 sessions, 391 prompts, 40 days, across three projects. Then I sorted every capability into two columns. Things I typed out loud and asked for. Things Claude chose on my behalf.

Look at the third row. I asked for browser control six times, so I know it exists and I know what it is called. Claude used it 1,295 times. Knowing the name is the entire difference between those two numbers, and it is the difference between using a tool and being handed results.

You cannot ask for what you do not know exists. Documentation does not fix this, because search only works for people who already have the word.

The shape of it

Understood versus executed

The teal band is three in all three columns. Everything that grew, grew above it.

The useful distinction

Only some of it is worth learning

Of the twenty steps you did not see, about five do not matter. Retries, file writes, glue. A machine should own those forever and you should never think about them again.

The other fifteen are different: plan mode, subagents, background tasks, hooks, MCP servers, artifacts. These are evergreen. They will still matter next year, in another project, with another model.

Here is the test I use. Would knowing this change what I ask for next month?

Claude retried a failed write: knowing that changes nothing. Claude ran five subagents in parallel across my refactor: if I knew that existed, I would ask for it the next time I had ten files to change. The first is plumbing. The second is a capability I am missing.

Knowing the name is what lets you ask for the bigger thing.

What we built

A tutor that watches, waits, and says one sentence

MONK's tutor mode keeps those evergreen capabilities in a library of fifty, watches what you are actually doing in Claude Code, and matches your current work against them.

When you write “add zod validation to all fourteen API route files, then typecheck each”, that is fourteen independent units of work. So it says: Claude has used agent teams for you forty times, and you have never asked for it once. Want to try it here?

The same sentence, sent while you were writing documentation, would be noise. Timing is most of the job. Silence is the default: most prompts get nothing at all, the card never steals focus, and one click dismisses it for good.

And accepting the suggestion is not the win. The win is four weeks later, when you type “run this in the background while I keep working” without being prompted, in your own words, because the concept is now yours.