Run /insights, Then Let Your Most Expensive Model Grade You

Most of us spend our best AI model on the wrong thing.

We save the big, expensive model for the hard code — the gnarly refactor, the tricky bug. Fair enough. But the single highest-leverage thing a top-tier model can do isn't write one more function. It's look at how you work and tell you where you're wasting your life.

There's a two-command habit for exactly this, and it takes about five minutes. I run it roughly once a month now.

Step one: ask Claude to report on you

Claude Code has a command most people have never run: /insights.

It generates a report analyzing your recent Claude Code sessions — the areas of your projects you spend the most time in, your interaction patterns, and the points where things kept getting stuck. It's reading your own history back to you. Not "here's how the tool works," but "here's how you've actually been using it."

That alone is quietly useful. Seeing your own patterns written down is a bit like hearing a recording of your own voice — slightly uncomfortable, and instantly clarifying. You notice you keep re-explaining the same context. You notice a whole category of work that's eating your afternoons. You notice the friction you'd stopped consciously feeling.

But the report is just raw material. The good part is what you do with it next.

Step two: hand it to your most expensive model

Once the report is sitting there in the conversation, switch to the best model you have access to. Right now that's Fable 5, the most capable model in the lineup:

/model fable

Then ask it, in plain English, to tear the report apart:

Critique this report on how I've been working. Where am I wasting time or repeating myself? What Claude Code skills, subagents, hooks, or slash commands would remove that friction? What should I change about how I work?

Now the expensive model is doing the thing it's genuinely best at — not pattern-matching syntax, but reasoning about a messy situation and giving you judgment. It reads your patterns, spots the repeated manual steps, and hands you a concrete list: this recurring task should be a skill; this thing you keep re-explaining belongs in a CLAUDE.md rule or a dedicated subagent; this friction point is a five-line hook away from disappearing.

Why spend the pricey model here, of all places

Because this is the one place the spend actually compounds.

A better model writing your next function gets you a marginally better diff — nice, but it's one diff. A better model reflecting on hundreds of your sessions can restructure the way you work for months. Cheap and fast where the work is routine; your very best model where the judgment pays off again and again. This is one of those spots.

And you run it rarely. This isn't your day-to-day driver burning premium tokens on every keystroke. It's a monthly ten-minute audit. The cost is a rounding error; the leverage is not.

Turn the critique into a backlog

Don't try to act on everything at once. Treat Fable 5's critique as a to-do list and pick one or two items:

• If it suggests a skill, run /skills first to see what already exists — Claude ships with more than you'd think, and half of what you "need" may already be there. Ask it directly: what skills could help my workflow?
• If it suggests a repeatable automation, that's usually a hook or a small slash command.
• If it suggests capturing context you keep retyping, that's a line in CLAUDE.md.

Implement a couple. Then next month, run /insights again and see if the friction actually moved. That loop — measure, critique, change one thing, re-measure — is the whole game. It's the same discipline we bring to shipping software, just pointed back at ourselves.

A few honest caveats

• The report is a starting point, not a verdict. /insights summarizes your sessions; it doesn't know your intent. Read it as a mirror, not a scorecard.
• Don't blindly install everything Fable 5 recommends. A model asked to suggest improvements will always find some. Sanity-check each suggestion against /skills and your own judgment before you go adding machinery you'll never use. More tooling isn't the goal — less friction is.
• Switch the model back when you're done. Once you've got the critique, /model back to your usual driver so you're not paying premium rates for routine work.

The reflex is to point your most powerful model at your hardest code. Try pointing it at yourself instead. Its most valuable output might not be a better function — it might be a better you, sitting at the keyboard, wasting a little less of every day.