Automate Personal Busy Work

Section 01

What this skill does

Automate Personal Busy Work is a skill that interviews you about your working week, works out which five recurring tasks are costing you the most, and then writes the machinery that does them. You install it, say "automate my busy work", and about half an hour later you have a folder containing five written procedures, five ready-to-paste prompts, and a plan for putting them to work one week at a time.

It is built for anyone whose week contains repetitive work they own personally. Freelancers and consultants, people running a small business, and people inside a bigger organisation with a repetitive slice of their job. It needs nothing connected to run. If you do have your email, calendar, or files connected, the automations get shorter, and the skill tells you exactly which connection removes which step.

Here is what you walk away with:

A ranked list of every recurring task in your week, with an honest estimate of what each one costs you per month

Five finished automations: an SOP, a prompt for your AI, a portable prompt that works anywhere, and either a scheduled task, an installable mini-skill, or a designed trigger

A list of the work it thinks you should not automate, with the reason for each

A list of the work you should probably delete instead

A four-week plan for adopting it, because five new habits started on one Monday do not survive until Tuesday

💡

The one rule the skill runs on: every session has to end with files you can use. If a run finishes with "you could probably automate your invoicing", it has failed.

Why it interviews you instead of just asking

The obvious way to build something like this would be to ask you which parts of your job you want handled. That approach does not work, and it is worth understanding why before your first run.

Try it right now. Ask yourself what you would automate if you could. You will come up with three things: email, invoices, and whatever irritated you most this morning. That is the standard answer, and it is almost always the wrong list, for three reasons.

Small tasks are invisible. Nobody reports the four-minute job they do eleven times a week. But four minutes, eleven times, is roughly three hours a month. That beats the ninety-minute job you do once a month, and yet the ninety-minute job is the one you will name first, every time.

Habitual work stops registering as work. If you have done it every Friday for two years, it feels like part of the furniture. It does not feel like a task with a time cost attached, so it never makes the list.

You report the job you think you have. Most people describe their role, not their week. The gap between the two is usually admin, chasing, and explaining the same thing to a different person for the fourth time.

So the skill does not ask the direct question. It rebuilds your week from evidence, ranks what it finds using actual arithmetic, and then builds the top five. The rest of this article walks through how each of those parts works, and where the skill falls short.

Section 02

What happens when you run it

A full run has six stages and takes about half an hour, most of which is you talking. Here is the whole shape of it before we go into the parts that matter.

1. Setup, once

Six questions about your role, your tools, what your AI can do, and what is off-limits. Saved, so you never answer them again.

2. The interview

Ten minutes, three passes. Rebuilds your week backwards, then triggers the work you have stopped noticing.

3. Scoring

Every task scored on time, repeatability, judgement, reachable inputs, annoyance, and risk. Ranked by arithmetic, not vibes.

4. The verdict

You see the ranked table before anything gets built, including what it thinks you should not automate and what you should delete.

5. The build

Five tasks, built one at a time. SOP, prompts, and either a scheduled task, an installable skill, or a designed trigger.

6. The rollout

A four-week plan for adopting them, because five new habits started on one Monday do not survive until Tuesday.

You can also skip most of that. Say "just automate this one thing" and it scores your one task and builds it. Say "re-run my busy work audit" three months later and it picks up from what stuck rather than starting over.

Section 03

The interview runs backwards on purpose

The interview is the part most people expect to be boring and most people end up finding useful on its own, before anything gets built. It never asks the direct question. It works in three passes.

Pass one: rewind the week

You go backwards from yesterday, one day at a time. Backwards works better than forwards. Recent memory is sharper, and it stops you narrating an idealised Monday-to-Friday that never happened.

For each day the question is not "what did you do" but "what was on that day, and what did you have to do around it". The "around it" is the whole point. A meeting is not a task. Preparing for it, writing it up, chasing the two actions it created, and updating the three people who were not there are four tasks, and those are the ones that repeat.

Pass two: trigger the invisible work

This is where the good candidates come from. A handful of questions, chosen because they retrieve work that direct questions never surface:

What did you do more than once last week?

What did someone have to chase you for?

What did you do while half-watching something else?

If you vanished for two weeks, what would be waiting in a pile?

What did you copy from one place to another?

What do you still do because someone asked once, and nobody has asked since?

If you dry up, it switches from asking to offering. It reads you four or five patterns common to your kind of work and asks which are yours. Recognition is far easier than recall, and most people add three or four tasks at this point that they would never have volunteered.

Pass three: numbers, in one batch

Here is the design decision that keeps this to ten minutes. It does not ask "and how often do you do that, and how long does it take" fifteen times over. That is slow, and your estimates get worse as you get tired.

Instead it fills in its own guesses for every task, shows you the whole table at once, and says: correct the rows I got wrong, ignore the rest. People edit far faster and more accurately than they estimate from scratch. You will change three or four rows and wave the rest through.

⚠️

Worth knowing: when it totals the table and reads the monthly number back to you, it is almost always bigger than you expected. That number is an estimate built from memory, not a measurement, and the skill labels it that way. It is still usually the moment people decide to keep going.

Section 04

How it decides what is worth automating

Every captured task gets scored on five things, and gated by a sixth. This is arithmetic, not a judgement call, which is deliberate: eyeballing a ranking gives you back the same top five you would have guessed on your own, and that is worth nothing.

What it measuresWhy it is in there

TimeTimes per month multiplied by minutes per go. This is where the counter-intuitive results come from.

RepeatabilityDo the steps stay the same and only the inputs change? Scored on the steps, not the subject matter.

JudgementHow much the outcome depends on a human reading a person, a room, or a context an AI cannot see.

Input reachCan the information actually be got at, given what you have connected? Not what is theoretically connectable.

AnnoyanceHow much you hate it. Deliberately kept in. A task that drains you costs more than its minutes.

RiskWhat happens if it goes wrong. Adds no points at all. It caps how far the task can go.

The time score is the one that produces surprises. Because it is frequency multiplied by duration, a four-minute job done daily outscores a ninety-minute job done monthly. That single piece of arithmetic is usually enough to reorder someone's entire sense of where their week goes.

Risk works differently from the rest. It is scored on what happens if a mistake goes uncaught, not on how likely a mistake is, and it never adds points. It only sets a ceiling.

Section 05

The four tiers, and why Tier A is rare

Scores become tiers, and the tier decides what gets built.

TierWhat it meansWhat you get

ARuns on its own. Output stays with you or inside your own systems.SOP, prompt, scheduled task, installable skill

BThe AI drafts, you approve and send.SOP, prompt, skill, plus a trigger or schedule

CYou still do it, faster.SOP, a fill-in template, prompt

DDo not automate this.Named, with the reason, and a call on whether to delegate, delete, or keep

Two things about this table are worth understanding before your first run, because both of them will look like the skill under-delivering if you do not know they are intentional.

Tier A is going to be rare for you. If your work is client-facing, which describes most people who sell a service, almost everything lands in B. Anything that leaves your hands and reaches another person gets a review gate, whatever it scored. That leaves Tier A for internal work: moving your own data around, filing and renaming, assembling your own morning brief, keeping your own lists current.

The skill states this plainly rather than promoting a task to fill the tier. The honest version of the offer is that none of these run without you looking at them, because they all reach a client. What they do instead is arrive finished, so looking takes two minutes instead of forty.

Tier D is not a failure state. A task that depends on a relationship, a negotiation, or reading a person gets worse when you automate it, not faster. Saying so is often the most valuable output of the audit. Ranking a weak candidate into the top five to fill the quota is the fastest way to lose your trust on the first run, so it does not do that.

🎯

It also hunts for things to delete. A report nobody has opened in four months. A check that has never once found a problem. The same question you answer live every week that could be answered once, in writing, where the asker can find it. Deleting a task returns all of its time. The best automation is the one that turns out not to be needed.

Section 06

What actually lands in your folder

The output is a folder called an Automation Pack. One index file and five task files. The index carries the ranked table, the hours targeted, the four-week plan, the do-not-automate list, and the delete list.

Each task file has five parts.

1. The SOP. What the task is, what starts it, what it needs, the steps, the decision rules, what "done" looks like, and how it goes wrong. Written to a specific test: could a competent stranger execute it on their first day without asking you a question? This is the part that outlives everything else. Prompts get rewritten when tools change. The SOP is what you hand to an assistant, a contractor, or yourself in eleven months.

2. A prompt for your AI. Written against what your setup can actually reach, referring to your tools by category rather than assuming a connection you do not have.

3. A portable prompt. Self-contained, with its own context block and everything pasted in. Works in any AI, on your phone, or in the hands of someone you forward it to. If you have nothing connected, the two versions collapse into one and it ships a single prompt instead of two near-identical blocks.

4. The automation, or the trigger. A scheduled task if your setup can run one. A small installable skill if your setup can install them. And if it can do neither, a designed trigger instead, which is the part covered next.

5. First-run checks. What to verify the first time, and the exact tuning lines to paste when the output comes back too long, too formal, too generic, or the wrong shape.

The trigger problem, and why it gets its own section

Plenty of people run their AI in a plain chat window with nothing connected. For them, the prompt is never the hard part. The hard part is that nobody remembers to run it. An automation nobody triggers is a document.

So when scheduling is not available, the skill designs the trigger explicitly. Usually one of four patterns: a recurring calendar entry whose title is the instruction and whose notes hold the prompt, an anchor onto something you already never miss, a standing file you open at a fixed point in the week, or a queue you check on a cadence.

Example trigger

Title: Run the Friday numbers thing (prompt in the notes)

Repeats: every Friday, 15:30

Notes: [the full prompt, ready to copy]

That is not a reminder. It fires at the right moment, in a place you already look, and it removes the "where did I put that prompt" step that kills adoption in week two.

Section 07

The move most people miss: splitting a task in half

This is the single highest-value step in the whole process, and it happens before scoring, on every task.

Many tasks sit in the middle of the ranking because they are two tasks stapled together: a mechanical half and a judgement half. Scored as one blend, you get a mediocre answer for both halves and a boring Tier B.

Split them, and the picture changes completely:

Reviewing applications splits into filtering against stated criteria, which is mechanical, and deciding who to interview, which is not.

Handling a support request splits into classifying it and pulling the relevant history, which is mechanical, and deciding what to offer an upset customer, which is not.

Writing the monthly report splits into assembling the numbers and drafting, which is mechanical, and the judgement call in the recommendation, which is not.

Automate the mechanical half. Hand yourself the judgement half with the groundwork already done. Almost nobody arrives having done this to their own work, and it is usually where the real hours are hiding.

Section 08

What it does well

It finds work you forgot you do

The recall triggers surface the four-minute jobs and the habitual work that never make it onto a list you write yourself.

It ranks with arithmetic

Frequency times duration reorders your priorities in a way instinct does not. The small frequent task usually wins.

It ships artifacts

An SOP you can hand to a person, prompts you can paste today, and a trigger that fires. Not a strategy document.

It tells you when to stop

It will say a task should not be automated, or should be deleted outright, rather than building something to fill a quota.

It works with nothing connected

No integrations required. Connections make the automations shorter, and it tells you which one removes which step.

It plans the adoption

One automation a week for four weeks, ordered so the habit forms before the hard one arrives.

There is a quieter benefit that comes up often. Even when someone builds nothing at all, they finish the interview with a written picture of their week and a delete list. Several people find that more useful than the automations.

Section 09

Where it falls short

Every one of these is a real limit. None of them are hidden in the skill, and knowing them before your first run is the difference between a useful result and a disappointing one.

The numbers are estimates, not measurements

Frequency and duration come from your memory on a given afternoon. The skill labels them as estimates and says "about" and means it, but if you want precision you need a real time log for a week first, and that is a different exercise.

Checking work does not compress like writing work

Drafting an update, a reply, or a write-up can genuinely go from forty minutes to five. Reconciling figures, matching records, or verifying a filing cannot, because being right is the job and confirming is most of the cost. For that kind of task the honest promise is a shortlist of the twelve lines that need your eyes instead of the four hundred that do not. Useful, and not the same as "it does the reconciliation".

It cannot see your systems

If the information a task needs is locked in a system nothing can reach, or lives in someone's head, no prompt fixes that. The skill will say so and tell you what would have to change, which is usually one connection or one decision about where something is kept. But it will not pretend.

Two of five is a normal result

Of five automations delivered, expect two to stick and become permanent, two to get used sometimes, and one to be abandoned. The rollout plan assumes exactly that. If you were expecting five for five, you will read a good outcome as a failure.

The first run of each one will not be perfect

You will add a line or two of correction. That is tuning, not failure, and once a correction has worked twice you paste it into the prompt permanently. People who are told this treat it as setup and keep going. People who are not told it conclude after one imperfect output that the whole pack is broken.

It creates a little work of its own

Drafts to review, digests to read, outputs to file. Some of that is worth automating in turn, which is one of the things the follow-up review looks for. Scheduled tasks in particular accumulate, and six months of unreviewed schedules produces a stream of output nobody reads, which is a new busy work problem created by the solution to the old one.

⚠️

The hard boundaries. Nothing goes to a customer without you reading it. Nothing scheduled will commit money, agree a date, or accept terms. Payroll, contracts, legal filings, and regulated reporting are off the table for unattended running. And whatever you name as off-limits during setup stays off-limits regardless of how well it scores.

Section 10

Who it fits, and who it does not

It fits anyone whose week contains recurring work they own personally. Freelancers, consultants, people running a small business, people inside a larger organisation with a repetitive slice of their job. The skill adapts to whoever runs it, because everything specific to you lives in your saved setup rather than in the skill.

It fits less well in two situations. If your work is genuinely different every day, with no repeating shape at all, the interview will find fewer candidates and the ranking will be thin. And if the work you want handled belongs to a team rather than to you, you need process documentation and delegation, which is a related but different problem. The SOPs it produces are a decent starting point for that, but the audit is built around one person's week.

Seasonal and project-based work is handled, by the way. Instead of "last week" it asks about the last time you were in the busy part of the cycle, and about the shape of the cycle itself, because the recurring work in that kind of business lives at the boundaries.

Section 11

Getting a good first run

Four things make a noticeable difference.

Have your calendar open. Pass one is a memory exercise and your calendar is the cheat sheet. It halves the time and roughly doubles what you find.

Answer with the boring stuff. The instinct is to report the impressive parts of your week. The value is in the four-minute jobs you are slightly embarrassed to mention. Those are the ones with the arithmetic on their side.

Use your own words for things. If you call it "the Friday numbers thing", say that. The skill keeps your name for it all the way through, because that is the name you will recognise in a folder in six months, and it is the phrase you will actually type when you want to trigger it.

Say what is off-limits, properly. Setup asks once. Anything you name there is excluded whatever it scores, and it saves you an argument with a ranking later.

Two ways to start

For the full audit, open a chat and say "automate my busy work". It will ask the six setup questions, then start the interview.

If you would rather test it on something small first, say "just automate this one thing" and name the task. It scores that task, tells you honestly if it is a poor candidate, and builds what it safely can. That is the cheapest way to see the output quality before committing half an hour.

Three months later, or after any change in your tools or workload, say "re-run my busy work audit". That pass opens with what stuck and what did not, hunts new ground rather than repeating itself, and builds three rather than five, because you already have a live set running.

Section 12

One thing to think about before you run it

The most useful output of this skill is often not an automation. It is the delete list.

Somewhere in your week there is a report nobody opens, a check that has never caught anything, a status update that exists in two places, or a question you answer live every week that could be answered once in writing. Those survive because nobody has ever questioned them, and they will keep surviving unless someone looks.

So before you go looking for the fastest way to do everything you currently do, it is worth asking which of it needs doing at all. An automation gives you back most of a task's time. Deleting it gives you back all of it.

Then run the audit on what is left.