Start here · Lesson 1.5
The FIND Framework for Choosing What to Automate With AI
Choosing what to automate with AI by feel gives uneven results. One week you hand over something sensible, the next you hand over something that needs more fixing than doing it yourself. A short routine you repeat removes most of that. This post gives you one, called FIND, made of four steps you can run in about half an hour.
FIND is our own working name for the routine taught in Lesson 1.5, Find the jobs to hand to AI. It is a practical checklist, not a research model, and you should adjust it to your work.
FIND stands for Fetch, Inspect, Narrow, Delegate: gather your tasks, score them, choose one, and hand it over with a written brief.
The four letters at a glance
| Step | Question it answers | Output |
|---|---|---|
| F: Fetch | What do I actually do? | A list of 25 to 40 tasks |
| I: Inspect | Which repeat and how do they score? | Scores out of 9 |
| N: Narrow | Which one is safe and worth it? | One chosen job |
| D: Delegate | How do I hand it over? | A written brief and a first run |
F: Fetch your tasks
Pull your tasks from records, not memory. Memory remembers the big and the annoying, and forgets the small repeating work that makes up much of a week. Use:
- The last two weeks of sent email.
- Your calendar and the follow-ups each meeting creates.
- Your task list or notes app.
- Your chat messages and customer questions.
Write one line per task. Do not sort yet. The aim is volume: 25 to 40 lines is typical for a normal working fortnight, and most people find the list is longer than they expected.
I: Inspect and score
Now go through the list. Mark anything that repeats at least weekly or that you know will return. For those, score three things from 1 to 3:
- Frequency: monthly is 1, weekly is 2, daily or near daily is 3.
- Duration: under five minutes is 1, up to half an hour is 2, over half an hour is 3.
- Handoff ease: needs your judgement each time is 1, partly patterned is 2, you could explain it on one page is 3.
Add them up. Totals of 7 to 9 are strong, 5 to 6 are possible, and 4 or below should stay manual for now. The scores are rough on purpose. They are there to stop you choosing by annoyance and to give you something to compare.
N: Narrow to one
From your top scorers, remove anything that fails a safety check. Ask three questions:
- What if it is wrong and I miss it?If the answer is costly, pick a different job or keep a human check on the last step.
- Does it involve private data, money or legal wording?Start elsewhere, and check the privacy terms of any tool before sharing customer information.
- Can I check the result in a minute?If not, the checking may cost more than the task saves.
What is left is your shortlist. Choose one. Resist the urge to take two or three, because a single well-briefed job teaches you more than three half-finished ones. Write it as a sentence: "Every [time] I [do this] using [these inputs] and produce [this output]." If the sentence will not come out cleanly, the task is too fuzzy, so go back to the shortlist.
D: Delegate with a brief
Handing over is a writing job. Anthropic's prompting guidance describes treating the AI like a capable new employee who lacks context on your norms and workflows, and says the more precisely you explain what you want, the better the result. It also lists examples among the most reliable ways to steer format, tone and structure. So your brief has five parts:
- The job: one sentence on what and why.
- The inputs: what it receives each run.
- What good looks like: paste a real example you are happy with.
- What to avoid: claims, phrases or topics that are off limits.
- Questions first: ask it to check anything unclear before starting.
Run it on material from last week where you already know the right answer. Compare, then fix the brief rather than patching the output by hand. More on writing a good one is in How to Brief AI Like a Senior Hire.
A worked example (illustrative)
This is an example, not a real case. Say you run a small photography studio.
Fetch: you scan sent email and find replies to enquiries, shoot confirmations, "your gallery is ready" notes, invoice reminders and testimonial requests.
Inspect: the gallery-ready note scores 3 on frequency, 1 on duration and 3 on handoff, a total of 7. The enquiry reply scores 3, 2 and 2, a total of 7. The invoice reminder scores 2, 1 and 3, a total of 6.
Narrow: the enquiry reply involves prices and promises, so you want to read it before it goes. The gallery note is low risk and quick to check. You choose the gallery note, since it is the safest of the high scorers.
Delegate: you write a brief with your best past gallery note as the example, tell it to avoid promises about edits you have not done, and ask it to check for the client's name and gallery link first. You run it on last week's three galleries and compare.
Whether this saves you time is something you learn only by running it. The point of FIND is that you test one job properly before moving on.
Running FIND every week
The first pass takes longest. After that, a weekly pass takes around fifteen minutes:
- Monday or Friday, pick a fixed slot. Fifteen minutes is enough.
- Add new tasks you noticed during the week to your list.
- Re-score anything that changed. A task that grew more frequent may now score higher.
- Check the live jobs. Is each hand-off still working and still checked?
- Choose the next job only if the last one is stable.
If you use a tool that works on files on your own computer, such as Claude Code, which its documentation describes as able to edit files and run commands, the same routine applies. The only difference is that the output may be a saved file.
When FIND says no
Sometimes your list has nothing above 6. That is a useful answer. It may mean your week is mostly judgement work, or that your tasks are too varied to brief. Do not force it. Try the "second time" habit instead: every time you do something for a second time, note it. In a month, your list will look different.
Questions people ask
Is FIND an established framework?
No. It is a simple routine we use to teach Lesson 1.5. Adapt it freely.
How often should I run it?
Once fully, then a short pass weekly or every two weeks.
What if two jobs score the same?
Pick the safer one, meaning the one where mistakes are cheaper and easier to spot.
Do I need special tools?
No. A notes app or a spreadsheet is enough to Fetch and Inspect. Any AI tool you already use can take the brief.
Where to go next
Run one full pass of FIND this week. For a guided walk through it, Lesson 1.5 covers each step: Find the jobs to hand to AI. If you want more ideas for what to look for, see 12 examples you can adapt.