The nine-step disaster
You ask an assistant to analyse six months of sales data, find the patterns, and build a presentation for the leadership team. Four minutes later you have twelve slides. They look wonderful. Then you notice slide nine has a number you do not recognise, trace it back, and discover that at step two it decided your regions were split east and west when in fact they are split by product line.
Everything after step two is now wrong. Beautifully formatted, internally consistent, and wrong. You cannot fix slide nine. You have to start again.
This is the single most common way big AI jobs fail, and the fix costs about ninety seconds.
In plain English
- Checkpoint:
- A deliberate stop where the work pauses for you to look before it continues.
- Plan first:
- Getting the list of intended steps before any of them run.
- Compounding error:
- A wrong turn early that every later step quietly builds on.
Ask for the plan before the work
The habit is simple. For anything with more than two or three moving parts, do not ask for the output. Ask for the plan.
A plan takes seconds to produce and about twenty seconds to read. In that twenty seconds you will catch the region split, the missing month, the audience misunderstanding, the fact that it was going to include the pilot data you wanted excluded. Catching it there costs nothing. Catching it on slide nine costs the whole afternoon.
I want you to do this job: [describe the whole thing in a few lines].
Do not start yet. First give me your plan as a numbered list of steps.
For each step say what you will produce and what you will assume.
Flag anything you are unsure about or any decision I should make.
Then stop and wait for me to approve or change the plan.
โ Weak prompt
Prompt
Analyse this sales data and build me a leadership presentation.
Output
Twelve finished slides, confidently laid out, built on a regional split the assistant chose for itself in the first thirty seconds.
One wrong assumption at the start, invisible until slide nine, and every later slide inherits it. The only repair is to start over.
โ Good prompt
Prompt
Analyse this sales data and build a leadership presentation. Do not start yet. Give me your plan as numbered steps, say what each step produces, and flag every assumption you are making about the data. Then stop.
Output
Six numbered steps, with a flagged assumption that regions are east and west, plus a question about whether to include the pilot months.
Two corrections typed in fifteen seconds, and then the whole job runs correctly the first time.
Checkpoint
Plan, approve, then run. Twenty seconds of reading a plan prevents the compounding error that makes every later step wrong.
Where to put the stops
Not every step needs your eyes. Put a checkpoint anywhere a decision gets made that everything downstream depends on.
Good places to stop: after the structure is chosen but before anything is written. After the data is interpreted but before it is charted. After the first section, so you can confirm the tone before nine more arrive in the same wrong voice. After the facts are gathered but before they are polished into prose you will be reluctant to throw away.
That last one matters more than it sounds. Nobody wants to bin twelve beautiful slides. Check the ugly version, because you will actually be honest about it.
The plan is approved. Now work through it with these rules:
Do step 1 only, then stop and show me the result.
Wait for me to say continue before starting step 2.
At each stop, tell me in one line what you assumed and what you are about to do next.
If anything I approved turns out not to work once you start, stop and tell me
rather than working around it quietly.
Say "then stop" and "wait for me" explicitly. Assistants are eager to be helpful and will run the whole job if you leave any doubt. Those four words are the difference between a checkpoint and a suggestion.
Cutting the job into pieces
Some jobs are better handled as separate handovers rather than one long chain. The rule of thumb: if the output of one part is something you would want to keep on its own, make it its own step.
Building a quarterly report might become: pull the facts, then draft the narrative, then write the summary, then produce the slides. Four checkpoints, four things you can inspect, and if the narrative is wrong you have not lost the facts.
Break this job into the smallest number of steps where each step produces
something I could look at and judge on its own.
THE JOB: [describe it]
For each step give me:
- what it produces
- what it needs from the previous step
- what would tell me it has gone wrong
Aim for [4] to [6] steps. Do not do any of the work yet.
Stop there. Step [2] is wrong: [what is wrong and what it should be].
Redo step [2] with that correction.
Then tell me what else in the plan needs to change as a result,
before you carry on. Do not silently patch later steps.
When you correct something mid-job, always ask what else the correction affects. A fix applied to one step without checking the rest is how the same error survives in a new form.
The economics of it
People skip planning because it feels like extra work. It is the opposite. A plan is thirty seconds of reading. A wrong finished deliverable is thirty minutes of forensics followed by starting again, plus the risk that you do not spot it at all and send it.
The larger the job you are handing over, the more the plan is worth. For a two-line email, just ask for the email. For anything you would have blocked out an afternoon to do yourself, get the plan first. Every time.
๐ Quiz
Question 1 of 4Why is a wrong assumption at step two so damaging?