Local or online AI?
Run a useful first test.

What you will make

A short comparison record for a text task: one local result, one hosted result, and a clear account of which needed less work. This is a practical selection exercise, not a model benchmark or a claim that one deployment method always wins.

You need: a local chat application with a compatible model already installed, access to an online assistant, a timer and a plain-text document for your notes. If local installation is not appropriate for your machine, compare two hosted tools and record that change to the exercise.

LM Studio’s official app guide is one route into local chat. Check its hardware guidance before downloading a model. For a browser-based option, consult the ChatGPT web guide. Features and account limits depend on the tool and plan.

1. Make the task small enough to judge

Use the fictional notice below. It contains a date, a location, one accessibility detail and one deliberately missing fact. That missing fact matters: a helpful-sounding invention is still an error.

The Riverside Makers Club will hold an open studio on Saturday 17 October 2026, from 14:00 to 17:00, at 12 Mill Lane. Visitors can see printmaking and paper sculpture demonstrations. Entry is free. The workshop is on the ground floor with step-free access. Refreshment arrangements have not been confirmed. No booking is needed.

Because this is invented sample material, the exercise does not require uploading client files. Substitute your own approved example later, once you understand the service and its data settings.

2. Use the same instruction in both tools

Start a fresh conversation in each tool. Paste this instruction and the notice together. Leave optional web search and extra integrations out of this first comparison so the source material remains the same.

Using only the notice below, write a welcoming event announcement of 70–90 words. Include the date, time, address, activities, entry cost, booking requirement and accessibility information. Do not invent arrangements that are unconfirmed. Then list any missing information that an organiser should check before publication. Keep that checklist separate from the announcement.

This is an original practice prompt, not a tested performance claim. The task structure follows a useful general principle: state the goal, provide context and specify the output. OpenAI’s prompting guide explains those elements.

Record the application, selected model, date and visible settings. Also note where processing happens. A desktop window can connect to a remote model; the window is not a map of the network.

3. Check the output before improving it

Read each result against the original notice. Do not ask the same system whether it did a wonderful job and count its enthusiasm as evidence.

• Facts: are the date, hours and address unchanged?
• Coverage: are both activities, free entry, no booking and step-free access included?
• Restraint: does it avoid promising refreshments?
• Format: is the announcement within the requested word range, with a separate checklist?
• Usability: does it sound like a real notice rather than an advertisement for the invention of Saturday?

Mark each requirement as met, partly met or missed. Save the untouched output before editing it. This gives you evidence of the first attempt rather than a memory improved by subsequent success.

4. Count the work you still have to do

Measure the time spent reviewing and correcting the announcement. Record generation time separately from initial setup. A model download is a setup cost; fixing the date is a task cost. Mixing them together obscures both.

Give both tools the same follow-up if they miss a requirement: “Revise the announcement to satisfy the original brief. Check every supplied fact and keep the checklist separate.” Record whether the revision fixes the problem or creates another one.

Repeat the exercise with a second notice: a changed opening time, a cancelled activity or an unknown address. One neat result is enough to continue exploring. It is not enough to trust a workflow with every notice you publish.

5. Write a decision you can revisit

Finish your record with these fields: task; tool and model; deployment; requirements missed; correction time; export friction; setup effort; next test. Avoid reducing everything to a single impressive-looking score.

You might choose local processing for a particular information boundary, a hosted tool for convenient collaboration, or neither because the existing template already does the job. A useful comparison permits all three answers.

The winning tool is the one that helps you finish this work under your constraints. Keep the record. Next month’s shiny announcement should have to compete with it.