One Prompt, Three Jobs. Three Mediocre Outputs.

TL;DR

Bundling three tasks into one prompt doesn't save tokens — it dilutes focus. The agent optimizes for the first objective; subsequent tasks get the remaining context. The result: three mediocre outputs instead of one good one. Split prompts when quality matters more than convenience — and learn to recognize when bundling is actually fine.

One afternoon I asked an agent to do three things in one shot: analyze a customer dataset, draft a summary email, and pull out key points for a presentation.

All three came back. The analysis had depth but missed the angle I needed. The email covered the facts but had the wrong register. The summary turned out to be a rehash of the analysis — not the actual problem I needed to present.

I spent an extra hour redoing each piece.

Next time I ran three separate prompts. Each came out better. Total time was less than before.

01Why Context Gets Diluted

When you put multiple objectives into one prompt, the agent doesn't process them in parallel. It works through the prompt sequentially — top to bottom. The first objective gets full attention. The second gets whatever context and capacity remain after the first. The third gets what's left of what's left.

Then there's the context contamination problem: if the first task is data analysis, the analytical frame and vocabulary bleed into how the agent handles the second task. You wanted the email to sound positive and forward-looking — but it just finished processing a breakdown of declining numbers. The two contexts conflict, neither wins, and the output lands somewhere in between.

And there's something subtler: when you bundle tasks, the agent decides how much is enough for each before moving on. That call belongs to the agent, not you.

✓ Three separate prompts

Each prompt has full context and a single objective

▼

Agent optimizes rather than balances — each task gets full effort

▼

Each output reviewable independently: clear signal on what's working

✕ Three tasks, one prompt

Context is shared; later tasks get the remainder

▼

Agent balances across objectives — none gets full effort

▼

Bundled output is hard to diagnose: unclear which part suffered and why

The paradox: bundling three tasks to save time often costs more — because of the rework round on the parts that came out weak. Three separate prompts are usually faster than one bundled prompt plus rework.

02Signs You're Bundling Wrong

Bad bundling isn't always obvious when you first receive the output. Sometimes you only realize when you go to use it and something feels thin. Four signals:

When task bundling is hurting quality

The last output is noticeably thinner than the firstthe first task got full attention; the last got the remainder — uneven treatment in a single prompt

Tone or framing shifts inconsistently across sectionsthe agent starts in one register, moves to the next task, and drags the previous frame along — two sections that don't cohere, with no obvious cause

One task in the bundle gets misread or partially droppedtoo many objectives means the agent sets its own priority order — that order doesn't always match yours

You can't tell where the problem is when you reviewbundled outputs are hard to debug — unclear whether the issue is the prompt, a specific task, or the sequence

The last one matters most for learning: a single-task prompt gives you clean signal — "this type of task doesn't work well." A three-task prompt gives you vague signal — "something's off" — with nothing clear to pin it on or improve.

03When Bundling Is Fine — When It Isn't

Not every bundle is a mistake. There's a reasonably clear line.

Bundling works when the tasks share the same context naturally and the outputs don't need individual precision. For example: "summarize this piece and pull out three key points" — two steps, but the context overlaps completely and the second follows naturally from the first. Or: "write this briefly, in an informal tone, aimed at developers" — that's one task with multiple constraints, not three separate tasks.

Split when each task needs a different register, a different depth of analysis, or when the output will be used independently. If you need a good analysis and a good email and a good summary — not a decent-analysis plus decent-email plus decent-summary — split them out.

The simplest test: if one of the three parts comes back weak, do I have to redo all three? If yes — they're not really one job. Split them so they can fail independently, and fix only what actually failed.

One prompt, one job doesn't mean more typing. It means that when quality matters, you control exactly how much focus goes into each piece. That, counterintuitively, is usually faster than one bundled round followed by rework.