Most jobs are not worth handing over
The failure people expect from Cowork is that it will do a job badly. The failure that actually happens is that they hand it a job that was never worth handing over. It does the thing competently, and the thing turns out not to have needed doing, or to have needed doing by a person who could be held responsible for it.
So before any of the briefing skills matter, you need a way to sort your work into two piles. This lesson is that sort.
In plain English
- Blast radius:
- How far the damage spreads if the work is wrong. Who sees it, what it triggers, how hard it is to undo.
- Draft-tolerant:
- Work where a rough first version is genuinely useful, because you were going to edit it anyway.
- Multi-step:
- A job with several connected stages, where each one depends on what the last one found.
- Accountability:
- Someone being answerable for the result. It cannot be delegated to software.
Four tests for a good candidate
Run a job through these four. It needs to pass all of them, not three.
Is it multi-step? If it is one action you could do in a single exchange, a chat window is faster and you get to watch. Cowork earns its keep when the job involves going and finding things, then doing something with what it found, then doing something else with that. One step is a question. Six steps is a job.
Is it repetitive? The effort of writing a proper brief only pays back if you are going to reuse the brief. A job you will do once is often quicker to do yourself. A job you do every month is worth an hour of specification.
Is it draft-tolerant? Ask whether a version that is mostly right and needs your editing is genuinely useful to you. For a first pass at a summary, yes. For a figure you are about to put in front of a board, no, because mostly right is the same as wrong.
Is the blast radius small? If this went wrong in the worst plausible way, what are you undoing afterwards, and how long does it take?
The line you already know
Our agents course draws a line between read access and write access, and that line applies here without modification. Reading the wrong folder is awkward. Sending the wrong thing to a client is a conversation you have on Monday. Rather than re-deriving it, borrow the habit from that course wholesale: before you hand over a job, write one sentence describing the worst plausible version of it, and see how you feel about that sentence.
Cowork's approval step is the structural answer to this. It is also not a substitute for the sort, because an approval you give without real understanding protects nobody.
โ Weak prompt
Prompt
Go through the outstanding invoices, work out who is late, and chase them.
Output
Eleven chasing messages leave, including two to clients who paid last week and one to a customer already in a payment dispute.
Money, external recipients, and messages that cannot be recalled. The blast radius is other people's inboxes and your credit control relationships.
โ Good prompt
Prompt
Go through the outstanding invoices and produce a chase list. For each item more than 7 days overdue: who owes it, what it is, how many days late, and a two line message I could send. Order by most overdue first. Flag anything that looks like it may already be paid or disputed. Draft only. Do not contact anyone.
Output
A ranked table of nine items, two of them flagged as possibly settled, with a draft message beside each.
Identical work, all of the tedium removed, and the irreversible part still belongs to you. This is the same job with the blast radius cut to zero.
Checkpoint
Multi-step, repetitive, draft-tolerant, small blast radius. All four, or keep it.
What to keep
Some work should stay with you regardless of how well it passes the four tests.
Anything with legal or financial consequence. Contracts, pricing, tax, anything that commits money or creates an obligation. Not because the output would be poor, but because the cost of a plausible-looking error is not proportionate to the time saved.
Anything where being confidently wrong is expensive. This is the one to watch. These systems produce a wrong date, a wrong name or a wrong figure in exactly the same steady voice they use for correct ones. There is no wobble to alert you. Where a confident error costs more than the whole task saved, the maths does not work.
Anything requiring accountability. Decisions about a person's job, a professional judgement your name sits under, anything a regulator or a client could ask you to justify. You can use it to prepare the material. The judgement stays yours, because accountability is not a thing software can hold.
HAND IT OVER when all four are true
[ ] several connected steps, not one
[ ] you will do it again next month
[ ] a rough version is genuinely useful to you
[ ] worst plausible outcome is undone in minutes
KEEP IT if any one of these is true
[ ] it commits money or creates a legal obligation
[ ] a confident error costs more than the task saves
[ ] someone could ask you to justify it
JOB Monthly supplier summary
STEPS read the folder, pull dates, compare against last month,
write the summary, list what changed
REPEATS first working day of every month
TOLERATES yes, I edit it before it goes anywhere
BLAST produces one draft document, sends nothing, changes nothing
VERDICT hand it over
Anything that sends, publishes or posts without me reading it first
Anything that quotes a price or agrees a deadline on my behalf
Anything touching pay, contracts, tax or legal terms
Anything about a specific person's performance or employment
Anything I would have to defend to a client, an auditor or a regulator
Anything where I could not tell a wrong answer from a right one
That last line is the sharpest test on the list. If you could not spot an error in the output, you cannot review it, and if you cannot review it the approval step is theatre. Handing over work you are unable to check is the one move that converts a safety feature into a liability.
๐ Quiz
Question 1 of 4Which four tests should a job pass before you hand it over?