Which tool to choose
- A long but text-only task (research, writing, step-by-step analysis): a capable assistant handles the stages within the same conversation, if you guide it one at a time.
- A task that touches several tools (search, write a file, send a message): agent mode, which plans and executes a sequence using external tools.
- A task that repeats with the same structure every time: it's worth turning it into a step-based automation on a no-code platform, so you don't re-guide it from scratch each time.
How to do it
The principle: a complex task is a chain of simple tasks. If one link breaks, the following ones inherit the error.
- Have it propose the breakdown first of all. The operational syntax:
I have to [complex objective]. Before starting, break it into the minimum necessary stages, in order. For each stage, tell me what it produces and what it needs from the previous stage. Don't execute yet: I'll wait for my go-ahead.
- Correct the plan. A missing stage or one in the wrong order costs a second here; discovered halfway through execution it costs all the work done.
- Execute one stage at a time, with a checkpoint. The operational syntax:
Execute only stage 1. Show me the result and stop. Don't move on to stage 2 until I confirm.
- Verify the result of each stage before giving the go-ahead to the next. The downstream stages trust the upstream ones: a wrong piece of data in stage 1 contaminates everything else without the AI noticing.
- Ask for a recap of what the next stage needs. When a conversation gets long, the AI can lose sight of the results of the first stages. Before continuing, take stock: "recap the results of stages 1 and 2 that we'll use now."
A concrete example
Sofia has to produce a market report. A big task. She asks for the breakdown: stage 1 gather the industry data, stage 2 organize it into a table, stage 3 identify the three main trends, stage 4 write the report. She approves the plan. She executes stage 1 and checks: the AI mixed data from two different years. She corrects it right away. Had she let everything run in one go, the final report would have been built on wrong numbers and she'd have discovered it only at the end, to be redone. By stopping at each stage, every brick is solid when the next one rests on it.
When it does NOT work (and how to fix it)
If the AI jumps ahead and executes all the stages together
You lose the checkpoints and end up with a final result where you don't know where it broke. Remedy: physically split it into separate messages. One message for the plan, then one message per stage, each saying "only this one, then stop."
If halfway through the AI forgets the results of the first stages
In long conversations the initial data "slips away" from the working memory. Remedy: before each new stage, paste back or have the needed results of the previous stages recapped. Keep a document with the outcomes yourself, so you don't depend on its memory.
If a stage produces an error the following ones don't notice
The downstream stages take what they receive as good. Remedy: add an explicit check to each stage ("check that the table's totals add up before continuing") and don't give the go-ahead until the check has passed.
A tip from someone who really uses it
Save the result of each stage in a document of your own, outside the chat. It becomes the project's reliable memory: if the conversation gets confused or you have to start over with another assistant, you restart from the good results instead of from scratch. The chain of stages holds up if every link is saved somewhere that doesn't depend on the AI's memory.
Frequently asked questions
How many stages are too many?
As many as it takes for each one to be checkable at a glance. If a stage is still too big to verify, split it. The limit isn't the number, it's your ability to say "this stage is right" before continuing.
Can I let the AI manage the stages on its own?
For small, low-risk tasks, yes. For tasks that matter, no: the human checkpoints are what stops an initial error from propagating. Total autonomy on a complex task is the surest way to get a convincingly wrong final result.
Is it true that a powerful AI can do everything in one shot without breaking it down?
It's the temptation that wastes the most time. Even the most capable assistant, on a multi-stage task done in one shot, accumulates small errors that become big at the end. Breaking it down isn't a limitation of the model, it's the way to be able to control it.