LLM Models Part 4: Find the Floor (Practice)

Michael Wacht

🔥

7d (edited) • ✨ AI Practitioner

In Part 4, we talked about dialing model capability up or down.

📌 Now let’s practice it.

My shorthand is:

Find the floor, then move back up one level.

That means you reduce the reasoning or effort level until you can see the model starting to miss the mark. Then you move back up to the lowest setting that still produces the quality you need.

That gives you a practical balance between:

Output quality
Speed
Cost
Available usage or capacity
The amount of correction you have to do yourself

Practice This:

Choose an assignment you already understand well.
Do not use something completely unfamiliar.
You need to be able to recognize when the output gets weaker.

Good practice assignments include:

A business proposal
A project plan
An app framework
A book chapter outline
A training lesson
A process improvement plan

✅Step 1:

Use the model map from Part 3 to choose your product and model.
Start with the model you believe is appropriate for the work.

✅Step 2:

Run the assignment using a relatively high reasoning or effort setting.
Your goal is to create the first strong version of the plan, framework, proposal, or structure.
Save the result.

This becomes your quality baseline.

✅Step 3:

Once the main thinking is complete, reduce the reasoning or effort level one step.

For example:

High reasoning → Medium reasoning

Or:

High effort → Medium effort

If the work is now very clear, and the product allows it, you can test reducing it two levels.

✅Step 4:

Ask the model to continue the work.

For example:

Expand one section
Rewrite part of the proposal
Create a summary
Draft an email from the plan
Turn the framework into a checklist
Create social posts from the article
Refine the tone
Format the output for easier reading

This is where the test begins.

You are no longer asking the model to do the hardest thinking from scratch. You are asking it to continue from an established structure.

✅Step 5:

Pay attention to what changes.

Look for:

Did it keep the original strategy intact?
Did it preserve the logic?
Did it maintain the theme or storyline?
Did it follow the framework?
Did it make weaker assumptions?
Did it require more correction?
Did it still make useful iterative changes?

You are not only asking whether the answer is “good.”

You are asking whether the model is still carrying the assignment properly.

✅Step 6:

Make one meaningful change that should affect several parts of the assignment.

For example:

The implementation timeline has changed from six months to 90 days. Update the plan accordingly.

Then watch what happens:

Does the model understand the broader impact of the change?

Or does it only patch the obvious section?

A strong result should recognize that a 90-day timeline may affect the recommendation, staffing, priorities, risks, budget, milestones, and next steps.

If the model only edits the paragraph you pointed to, the setting may be too low for the assignment.

That is the test.

🔑 The Takeaway:

If the model starts losing context, missing important logic, making weak assumptions, or requiring too much correction, dial up.

If the structure is already established and the work is predictable, creative, or execution-focused, dial it back.

Use enough model capability to do the assignment well. More is not automatically better.

11 comments

LLM Models Part 4: Find the Floor (Practice)