PRACTICAL FIELD NOTES / 18 OCTOBER 2026
How to choose an AI skill for your business
Choose a repeatable task with inputs you can supply and an output you can check.
Start with last week’s work
Pick a task you actually finished recently: preparing a proposal, explaining a reporting variance, organising an event, or answering a recurring customer question. Keep the original input and your reviewed output as the baseline. This gives you something concrete to compare with a skill-assisted draft.
Define the handover in one sentence. “A proposal with our approved prices and exclusions” is easier to evaluate than “automate sales”. Start with one bounded job rather than an autonomous department.
Check the setup before the promise
Read the publisher’s complete skill definition and supporting files. Does it use a CRM, command-line program, paid API or local script? Can your chosen app actually access those tools? A prompt cannot create a connection or permission that you have not configured.
Fieldwork labels sources and dependencies separately from execution testing. A source-inspected listing tells you which file we reviewed; it does not certify that the skill will work correctly in your account.
Put Australian assumptions in writing
State your country, state or territory, currency, reporting period and approved business policies. For broadly portable marketing or operational tasks, that may be enough to set context. For tax, employment, legal or clinical work, use the appropriate local professional and verify current primary sources before relying on any result.
Begin with the Australian-context collection when local rules matter. A worldwide label means a workflow travels reasonably well; it does not mean it implements every country’s regulation.
Measure the whole task
Use a permitted or fabricated sample first. Record the time spent preparing inputs, installing dependencies, reviewing the draft and correcting errors. Compare the final reviewed handover with your baseline, including completeness and accuracy. If a workflow saves drafting time but adds more review work, your measurement should show it.
Only repeat a workflow after the first result passes your usual checks. Save the version, prompt and context that produced it, so later changes can be assessed against the same example.