Pick one repeated task, collect examples of good work, and use AI to prepare the next result. Record every correction, turn those corrections into instructions, then test again. Compare total time and quality before deciding whether the task is worth automating.
The five-day method
Nick introduced this exercise in The AI Core as a way to start with a real job before buying or connecting more software.
Day 1: pick the repeated task
Choose a small job your team performs often. Make its starting input and finished output explicit.
Day 2: gather good examples
Collect five completed examples that show what acceptable work looks like.
Day 3: prepare the next result
Give AI the examples and the next input. Compare the output with the work your team needs.
Day 4: record the corrections
Turn recurring edits into a short set of instructions. Include what the system must leave for a person.
Day 5: try again
Repeat with those instructions. Check whether the output is useful and whether checking it saves time overall.
Keep a simple test log
A useful extension to the exercise is a small log of inputs, outputs and corrections. It prevents an impressive result from one easy example becoming the whole evaluation.
| Record | What to capture |
|---|---|
| Case | An internal reference to the input you tested |
| Manual time | How long the original job takes |
| Assisted time | Preparation, generation, checking and correction |
| Quality | Accepted, corrected or rejected, with the reason |
| Next instruction | A specific rule that would prevent the same error |
Example: preparing a routine reply
Imagine an operations team drafting replies to requests for missing documents. The first test can prepare a reply for someone to review; it does not need to send messages or change customer records.
Use one complete request, one with missing information and one that does not belong in this process. Check that the draft identifies the right document and leaves unknown details unresolved. An instruction to flag an unknown is more useful than a polished answer containing a guess.
This is an illustrative test scenario, not a reported client result.
What makes the task ready to automate?
Look for consistent quality across different inputs, a reviewer who understands the exceptions and a clear reduction in total work. Keep a human check wherever the pilot still depends on judgment.
If checking takes as long as writing the original, inspect the inputs and instructions. You may need a smaller task, better examples or a clearer definition of completion. The exercise is valuable even when the answer is to keep the job manual for now.
Sources behind this guide
Explore Nick’s original material and the Reprise approach this guide draws on.
- Newsletter · August 12, 2026Give AI one task for five days Nick Puru