Set up a model for each kind of task
Knowing which model suits a task is useless if switching is a trip into Settings. This is the setup that makes the right choice the easy one.
Why good intentions fail here
Everyone agrees in principle that a quick rewrite does not need a heavyweight model. Almost nobody acts on it, because acting on it means opening Settings, editing a field, saving, and remembering to change it back.
So the fix is not discipline, it is configuration. Make each choice already exist as a profile, and switching becomes one click in the composer.
Three profiles worth having
| Profile | Model | Reasoning effort | For |
|---|---|---|---|
| Fast | The smallest model that writes acceptably | Provider default | Rewriting, summarizing, translating, variants |
| Everyday | A mid-range general model | Provider default | Explaining, extracting, comparing, most work |
| Deep | Your most capable model | High, where supported | Debugging, design review, contradiction hunting |
Duplicate an existing profile and change the model rather than building each from scratch: the key is copied and the base URL is already right. Name them for what they are for, not for the model they contain, so the names survive changing models later.
Set Everyday as the default. Most work belongs there, and defaults are what actually get used.
A fourth profile worth considering
If you handle anything confidential, add a Local profile against Ollama or LM Studio.
The value is not only privacy but decisiveness. When *should I send this document to a hosted provider* is a question you have to think about each time, you will sometimes get it wrong under time pressure. When there is already a local profile one click away, the answer is just to switch.
Building the habit
- Glance at the composer picker before sending. It names the profile and model; that glance is the whole discipline.
- Default to Fast for anything you would describe as *just*. *Just rewrite this*, *just summarize this*.
- Escalate when an answer looks shallow rather than starting deep. Most problems that look hard are not.
- Switch to Local before attaching anything confidential, not after.
- Review the ledger after a fortnight. If Deep dominates your usage, either your work really is that hard or the habit has not formed.
Revisit it occasionally
Models change. A model that was your Deep choice last year may be this year's Everyday, at a fraction of the price. Re-run model discovery every few months and reconsider what sits in each slot.
Because the profiles are named for their role rather than their model, updating them is editing one field. That is the point of naming them that way.