Quick fixes AG-P004
One request, five tasks
Shared methods · A shared method; linked tool guides explain the exact steps.
This page covers tools outside your selection. You can still read it. Find matching guides
A prompt doing five jobs gets five mediocre answers stapled together. Split it, and each step gets the model's full attention plus your checkpoint.
The part I cared about got two sentences? Here's the quick fix ↓ — the why sits right below it.
Quick fix: One request, five tasks
Instead of Summarise this report, then write an email to the team about it, make a slide outline, and also check the numbers in section 3.
Try First: check the numbers in section 3 against the table on page 2. We'll do the summary after that's settled.
Do the step whose output the other steps depend on. Then the next one.
Why this works
Bundled tasks compete. The model does all five, none deeply, and an error in the early step (the numbers were wrong) silently poisons everything built on it in the same breath. Splitting is not slower — it front-loads the step you would otherwise redo everything for, and gives you a checkpoint between steps where steering costs one sentence instead of a full restart.
When this applies
The tell is the word "then", twice, in your own prompt, or an answer where the part you cared about most got two sentences.
Template
Step 1 of 3: [first task]. Just this — I'll ask for the next step after.
…and after checking the answer:
Good. Step 2: [next task], using what we just established.
Go deeper
The checkpoint habit is the core of prompts that survive a long task. For work that is genuinely one big job, when to start fresh instead of correcting covers the other half: knowing when a conversation is carrying too much. The multi-step version with different models per step is the workflows track.
Improve next
- Say what you actually want
Tell the AI what outcome you actually want.
- Make the next rewrite more useful
Say what was wrong with the last version before asking again.
Something wrong with this page?
Say what you expected and what you got. That is usually the shortest route to a correction, and it goes on the public issue tracker so the fix is visible.