Does a Longer Prompt Get Better Results? (Density vs Padding)

TL;DR: Length is not the variable. A longer prompt gets better results only when the extra words carry task-relevant information: constraints, format, audience, examples, source material. Words spent on emphasis, flattery, repetition or restated rules add cost without adding signal, and vendor guidance says repetition can actively destabilise the answer.

The claim that longer prompts work better is half-true, which is why it survives. People who write detailed prompts do get better answers. But the detail is doing the work and detail happens to be long, so the mechanism gets mistaken for the side effect and the advice mutates into "write more". That is where it stops working.

Does a longer prompt get better results?

Not on its own. Wherever a longer prompt beats a shorter one you can point at a specific piece of information the longer version added. Remove that piece but keep the length, and the advantage goes away.

The variable that moves the answer is information density: how much of what you typed narrows the space of acceptable outputs. A 40-word prompt naming the audience, the format and the one thing that must not appear is denser than a 300-word prompt built from three restatements of "be thorough and detailed."

This is also why the length question has no published answer. No major vendor publishes a recommended prompt length, and those publishing anything adjacent publish the opposite: guidance on what to remove. OpenAI's prompting guidance for GPT-5.6 opens its first working section with an instruction to simplify, stating that "Removing repeated instructions and examples and simplifying tool descriptions can improve task performance and token efficiency."

Prompt length is an output of good prompting, not an input to it.

What are the extra words actually doing?

Every sentence you add falls into one of two buckets, and sorting your prompt into them is the whole technique.

Payload is information the model does not have and cannot infer. Packaging is information about how you feel about the request, and the guidance below suggests that does not help.

The diagnostic comes from a vendor rather than from us. Anthropic's Claude Code best-practices page, on standing instruction files, says to ask of every line "Would removing this cause Claude to make mistakes?" and, if not, cut it. Run that on each sentence of any prompt. Payload survives. Packaging never does.

Which additions reliably help?

Anything that removes a decision the model would otherwise make for you, roughly in order of impact:

1. The output format. Prose, table, JSON, list, and the fields it must contain.
2. The audience. "For a CFO who has never used the product" beats three sentences of tone instructions.
3. Hard constraints. Length limits, banned words, what must not be recommended, the library version you are on.
4. Source material. Pasting the document beats describing it, every time.
5. A worked example. One few-shot example of the shape you want usually outperforms a paragraph describing that shape.
6. What "done" looks like. A success criterion tells the model when to stop, which is a failure mode in its own right.

Anthropic's best-practices page demonstrates this with its own before/after pair. Vague: "add tests for foo.py". Specific: "write a test for foo.py covering the edge case where the user is logged out. avoid mocks." Four times longer, and every added word is payload: which file, which scenario, which technique to avoid. The page's framing is "The more precise your instructions, the fewer corrections you'll need."

Here is the same transformation on a request people actually send. The useful version is longer, but not by padding.

BEFORE (vague, 9 words)
Write a blog post about our new pricing page.

AFTER (dense, 85 words)
Write a 600-word announcement post about a pricing change.

Audience: existing customers on a monthly plan who will see a
new annual option.
Facts: annual billing saves two months; monthly pricing is
unchanged; existing customers keep their current rate until
they choose to switch.
Format: 3 H2 sections, no bullet lists, no call to action in
the first 200 words.
Do not: describe the change as an increase, or imply the
monthly plan is being retired.
End with: one sentence naming where to switch plans.

And here is a padded version, set to almost exactly the same length as the dense one.

PADDED (84 words, almost no payload)
I really need your help with something IMPORTANT. Please write
an AMAZING, high-quality, engaging blog post about our new
pricing page. This is critical for our business so please take
your time and think very carefully about it. Make it really
good. Make sure it is well written and professional and
compelling. It is very important that it is high quality.
Please be thorough and detailed. I will tip you $200 for a
great job. Thank you so much, I really appreciate your help!

One word shorter than the dense version, and it still never says who is reading it, what changed, or how long the post should be. Count the words and the two are equal; count the decisions removed and one of them removed none.

Which additions do not help?

The folklore ones. Here a vendor has put something in writing, and it points against the technique rather than for it.

OpenAI's GPT-4.1 prompting guide, in its instruction-following section, says: "It’s generally not necessary to use all-caps or other incentives like bribes or tips. We recommend starting without these, and only reaching for these if necessary for your particular prompt." It then adds a warning most people miss: "Note that if your existing prompts include these techniques, it could cause GPT-4.1 to pay attention to it too strictly."

Read that last sentence again. The failure mode is not that emphasis is ignored. It is that emphasis works, on the wrong thing, dragging attention toward whatever you shouted.

The GPT-5.6 guidance arrives at the same place from the other direction, with a trim list: repeated statements of the same rule, repeated style instructions that do not change behavior, examples that do not change behavior, process instructions for behavior the model already performs reliably, and unrelated tools. It also rations absolutes: "Use ALWAYS, NEVER, must, and only for true invariants such as safety rules, required fields, or actions that should never happen."

OpenAI publishes a number alongside that advice, worth quoting with its own caveat attached. In a sample of internal coding-agent eval runs, it reports, "configurations with leaner system prompts improved evaluation scores by roughly 10–15% while reducing total tokens by 41–66% and cost by 33–67%." The next sentence matters as much: "Results will vary by workload, so treat these ranges as directional and validate changes on representative tasks from your own application." That is a vendor calling its own figure directional. Treat it that way, and note which way it points: shorter scored better.

Why do contradictions cost more than gaps?

Because a gap gets filled with a default, while a contradiction gets resolved by a rule you did not write.

OpenAI's GPT-5.6 guidance is unusually direct: "GPT-5-class models follow prompt contracts closely, so conflicting rules can create more instability than missing detail." Missing detail is a known cost. Conflicting detail is an unknown one.

The GPT-4.1 guide documents the resolution. Its debugging workflow says to "Check for conflicting, underspecified, or wrong instructions and examples", then gives the tiebreak: "If there are conflicting instructions, GPT-4.1 tends to follow the one closer to the end of the prompt." That is described as a tendency in one model family, not a law across all of them.

This is where padding turns from wasteful into harmful. Contradictions are rarely written on purpose; they accumulate. You add "keep it brief" in one paragraph and "cover every objection in detail" in another, or paste a style guide that disagrees with a rule three lines up. The longer the prompt, the more surface area for two of your own sentences to disagree, and the less likely you are to notice.

If you have wondered why AI ignores your format instructions, a competing instruction further down the prompt is the first thing to check.

Does a long instruction file make the model ignore instructions?

For standing instruction files, one vendor says yes, in a sentence ending with an exclamation mark.

Anthropic's Claude Code best-practices page, on keeping a project's CLAUDE.md short: "Bloated CLAUDE.md files cause Claude to ignore your actual instructions!" It gives the diagnostic too: "If Claude keeps doing something you don't want despite having a rule against it, the file is probably too long and the rule is getting lost."

Be precise about scope: that is guidance about a persistent instruction file loaded every session, not a study of chat-prompt length. But the mechanism is not specific to files. It is attention competing against volume, and a long chat prompt faces the same competition.

The page's line on emphasis is the neatest statement of the whole problem: "If you emphasize many lines, none of them stands out." Its fix is to emphasise one line only, the one the model keeps skipping. Emphasis works while it is scarce. Same for rules, examples and words.

That argument is developed at length in our post on what belongs in a CLAUDE.md.

Is "lost in the middle" real, or is it folklore?

A real, published finding, quoted far past what it established.

The source is "Lost in the Middle: How Language Models Use Long Contexts" by Nelson F. Liu and colleagues, on arXiv as 2307.03172, posted 6 July 2023 with a third version that November. The abstract reports that "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models."

Now the part that gets dropped. The paper names the models it tested: the open ones were "MPT-30B-Instruct, LongChat-13B (16K)" and the closed ones "OpenAI’s GPT-3.5-Turbo and Anthropic’s Claude-1.3". Those are 2023 systems. Every citation of the finding about a 2026 model is an extrapolation, whether or not the person citing it says so.

What does the extra length actually cost you?

Three things, and only the first is obvious.

Money. Input tokens are billed. A prompt three times longer costs roughly three times more to send, before the model produces a word.

Time. OpenAI's prompt caching guide lists reduced latency among the benefits of reusing a prefix: "Reduce the time spent processing input before the response starts." That processing is what a longer prompt buys more of.

Room. In a long conversation your prompt competes with every earlier turn for space in the context window. That is the link between prompt bloat and two common complaints: answers that cut off mid-sentence and a model that forgets what you told it.

One exception is worth knowing before you optimise. OpenAI's caching guide notes that a prefix below the minimum cacheable length can cost more than a longer one that qualifies for cached reads, concluding: "There is no universal maximum-cost prompt length; the crossover depends on reuse and pricing." For a stable system prompt sent thousands of times, longer can genuinely be cheaper. For a one-off question in a chat box, it is just longer.

How do you tell density from padding in your own prompt?

Run the subtraction test, then a real comparison if the answer still matters.

1. Delete every sentence about importance, urgency, effort or reward. These are the incentives OpenAI's guide recommends starting without.
2. Find every rule stated twice and delete one copy, keeping the more specific wording. Two copies are a future contradiction, because you will edit one and not the other.
3. Read what remains looking only for disagreements. GPT-5.6's guidance makes this its own review step, immediately after trimming.
4. Ask of each surviving sentence whether removing it would change the output. If not, cut it. Anthropic's line, and the only step that is hard to do honestly.
5. Now compare, properly. Same model, same settings, several runs of each version rather than one, graded against a rubric you wrote before seeing the outputs.

That last step is the one everybody skips, including most articles claiming to have tested prompt length. Two prompts and one output each is a coin flip described as a study. For more of the before/after habit, ten bad prompts fixed runs the same transformation across ten requests.

THE SUBTRACTION PASS — paste your prompt below this line and
run it as a single request.

Review the prompt below. Do not answer it. Do four things:

1. List every sentence that adds emphasis, urgency, stakes or
   praise but no task information. Quote each one.
2. List every rule that appears more than once, quoting both
   occurrences.
3. List every pair of instructions that could conflict, and
   say which one wins if the model can only satisfy one.
4. Rewrite the prompt keeping only sentences whose removal
   would change your answer. Report the before and after word
   counts.

PROMPT:
"""
[paste here]
"""

Where does a prompt tool fit into this?

On the "make it more specific" side, not the "make it longer" side. The two are easy to conflate.

Prompt Architects rewrites a prompt into a more specific one. The rewrite is usually longer than what you typed, for the same reason Anthropic's "write a test for foo.py covering the edge case where the user is logged out. avoid mocks." is longer than "add tests for foo.py": naming the format, the audience and the constraints takes words. Length is a side effect. If a rewrite is longer without being more specific it has failed, and you should cut it back. We have no measurement showing our output is better because it is longer, and we are not going to imply one.

The parts that matter here are the unglamorous ones. Saved prompts stop you rewriting from memory, which is where duplicated and contradictory rules come from. Variables and a personal context library keep the standing facts about your audience and product in one place instead of an approximation retyped into every prompt. Our FAQ states that "The Free plan includes 5 prompt enhancements per day, forever." Paid plans start at $4.99/month for Pro, which our pricing page lists at 200 enhancements per month, with unlimited on Advanced at $9.99/month and a 14-day money-back guarantee. Launch-promotion prices at the time of writing, checked 28 August 2026.

None of it makes the model care more because you asked nicely at greater length. Nothing does. The only lever is how much of what you typed the model could not have guessed, and the fastest way to add some is to delete something else first.