Don't Make One Prompt Do Five Jobs

The mistake of packing several distinct tasks into a single prompt — research this, then write it up, then edit for tone, then format as a table, and also suggest a title — and expecting all of them done well at once. Some parts land, others get skimmed or dropped.

A prompt with five bundled jobs forces Claude to split effort across all of them in one pass, and the usual result is uneven: the first ask gets real attention, later ones get token effort, and one quietly goes missing. Worse, when the write-up is good but the formatting is wrong, you have to re-run the whole thing to fix one part. Splitting the work into steps — each a single clear job — gives every step full attention and lets you verify and correct one link without disturbing the others. It's the difference between one overloaded turn and a short, controllable chain.

Notice this whenever your prompt contains 'and then… and also… and make sure to…' stacking several verbs. Also when an answer nails part of what you asked and ignores the rest — that's the signature of an overloaded prompt, not a careless model.

Genuinely small, tightly-coupled asks that belong together ('summarise this and give it a title') are fine in one prompt — the rule is against bundling distinct, separable jobs, not against ever asking for two related things at once.

Step 1 of a few — do only this, then stop and wait: <the first single job>.

(Then, in later turns: 'Now, using that result, do <the next single job>.')

Don't jump ahead to steps I haven't given you yet.

You ask Claude to 'research competitors, write a summary, turn it into a slide outline, and draft an email about it' in one go. The research is thin and the email is missing. Broken into steps — research first, review it, then summarise, then outline — each part is solid, and when the outline needs a tweak you fix just that step instead of re-running the entire request.

For tasks that genuinely have several stages, name the stages up front but ask Claude to do them one at a time with a checkpoint between each ('do step 1, show me, wait for go'). You keep the benefit of a plan while avoiding the overload — and you catch a wrong turn at step one instead of discovering it after all five ran.

• Reading a partial answer as Claude being lazy, when the prompt asked for more than any single pass can do well.
• Re-running the whole multi-part prompt to fix one weak section, instead of having split it into fixable steps.
• Chaining so finely that you split naturally-joined asks into needless turns — over-correcting the other way.