Guides

Guide · Workflows · 6 min read

Build Your Automation Backlog

Rank recurring work like a product backlog, pick one task, and run AI in shadow mode for a week before you hand it over.

Why this matters

It is easy to automate whatever you saw in a demo last night. A product team does not ship the shiniest ticket first. It ranks work by impact and risk, then starts at the top.

Your week is a backlog too. Score it once and the right first automation is often something boring that takes a few hours a month. That is exactly why it rarely shows up in a demo.

How to score a task

Score each recurring task on four things. The first two estimate the time it takes; the last two help decide how much to trust AI with it.

Four scores to estimate the value and risk of automating a task
ScoreWhat to write down
FrequencyHow many times you do it each month
MinutesHow long each run takes
Same steps?Yes, mostly, or no
Mistake costLow: fixed in a minute. Medium: someone notices. High: money, a client, or your reputation.

Hours per month = frequency × minutes, divided by 60. High hours, repeatable steps, and low mistake cost make a good first hand-off. If the mistake cost is high, let AI draft and review every result yourself.

The backlog prompt

Use this prompt to inventory the work you repeat, score it, and choose one task to test first.

Automation backlog · inventory and ranking
Help me build an automation backlog: a ranked list of my recurring work, scored like a product backlog, so I hand the right task to AI first instead of the flashiest one.

STEP 1: INVENTORY
Interview me one question at a time about the last two weeks. Walk through:
- My calendar: recurring meetings, and the prep or follow-up each one creates
- My inbox and messages: the kinds of replies I write over and over
- Documents: reports, updates, decks, or trackers I produce on a schedule
- Research and lookups I repeat
- Admin: scheduling, expenses, data entry, file cleanup
Keep going until we have at least 12 tasks. If I am vague ("emails"), split it into specific tasks ("answer vendor invoice questions").

STEP 2: SCORE EACH TASK
For every task, ask me for these, or estimate from what I already said and mark estimates with ~:
- Times per month
- Minutes each time
- Same steps every time? (yes / mostly / no)
- Cost of a mistake: low (I would fix it in a minute), medium (someone would notice), high (money, a client, or my reputation)
Then calculate hours per month for each task.

STEP 3: RANK THE BACKLOG
Show a table sorted by hours per month with these columns: task / times per month / minutes / hours per month / same steps? / mistake cost / mode.
Mode is one of:
- HAND OFF: same steps, low mistake cost
- DRAFT FOR ME: needs my judgment or carries a medium-to-high mistake cost, but AI can do the first 80%
- NOT YET: the steps change every time, or I am the point (relationships, final calls)

STEP 4: PICK ONE
Recommend the single task to start with, and explain why it beats numbers two and three. Then write:
1. The exact instructions I would save for it, ready to paste into a Claude Project, a ChatGPT Project, or a skill
2. What it needs access to, and the minimum access that would still work
3. A one-week shadow-mode plan: I keep doing the task myself, the AI does it in parallel, and we compare
4. The bar for full hand-off (for example: matched my version on 4 of 5 runs, with no high-cost mistakes)

RULES
- Do not invent tasks I did not mention.
- Show your hours math so I can check it.
- If nothing on the list is a good fit for AI yet, say so instead of forcing one.

Run it in shadow mode

Shadow mode means the AI does the task alongside you, but does not take action. Compare its work with yours until the results meet a bar you set in advance.

  1. Keep doing the task yourself for one week, exactly as you do now.
  2. Have AI do it too from the same input, without sending or changing anything.
  3. Compare each run with the prompt below. Add the most useful new rule to your instructions.
  4. Hand it off only after four of five runs match yours with no high-cost mistakes.
Shadow mode · compare one run
We are running shadow mode on this task: [task name]. I did it myself, and you did it in parallel from the same input. Compare the two honestly.

THE INPUT
[paste the email, notes, data, or request both versions started from]

MY VERSION
[paste]

YOUR VERSION
[paste]

SCORE THE RUN
1. Match: same outcome / minor differences / meaningfully different.
2. List every difference. For each one, say whose version is better and why.
3. Flag any mistake in your version that would have caused a problem if I had sent it unchecked. Rate each one low, medium, or high cost.
4. Write one new rule for your instructions that would have closed the biggest gap. Keep it to one sentence I can paste in.

TRACK THE TRIAL
This is run [#] of 5. Keep a running tally from earlier runs I paste below.
Hand-off is ready when 4 of 5 runs are "same outcome" or "minor differences" and none had a high-cost mistake. If we are not there after 5 runs, tell me whether to keep going, switch this task to DRAFT FOR ME, or drop it.

EARLIER RUNS
[paste previous scores, or "none yet"]

Do this next

Run the backlog prompt this week, but start shadow mode on only the top task. One automation you trust beats five you have to babysit.

Practical AI for founders & builders.

Get new guides, systems, and skills as they ship.