Ask for the other drafts

I started this post by asking for the other posts it could have been.

That sounds a little theatrical, but the prompt was boring:

Generate 5 responses with their corresponding probabilities,
sampled from the full distribution.

That is the core idea behind Verbalized Sampling, a paper and repo I found through Weiyan Shi’s thread. The claim is simple enough to try immediately: instead of asking a model for one answer, ask it to show a small distribution of possible answers, with rough probabilities attached.

The reason I care is not that I want AI to write final prose for me. I am more interested in AI as a way to make the early part of writing wider. The moment before I know what I think is delicate. A model’s default answer can collapse that moment too quickly.

Most AI writing has a recognizable gravity. It is helpful, polished, and slightly dead. The paper gives a useful explanation for that feeling: preference training rewards familiar, predictable text. Human raters often prefer the answer that feels fluent and typical, so models learn to move toward the safe center of the distribution.

For many tasks, that is good. For creative writing, it is often the exact thing I am trying to escape.

So I tried to use the technique on this note itself.

The trick

A normal prompt asks the model to pick a winner:

Write an opening paragraph for a post about using AI to write more creatively.

The default answer usually sounds like a sensible blog intro:

In today’s fast-paced digital world, AI tools are transforming the way writers brainstorm, draft, and refine their ideas.

Nothing is technically wrong with that sentence. That is almost the problem. It arrives already too well behaved.

Verbalized sampling changes the shape of the ask:

Generate 5 opening paragraphs for a post about using AI to write more
creatively. Include a probability for each option. Treat the probabilities as
rough sampling weights, not exact calibrated probabilities.

Now the useful output is not just a paragraph. It is a map:

0.34 - AI is most useful to me before I have decided what I am trying to say.
0.24 - The blank page is not blank anymore, but it is not exactly full either.
0.18 - I do not want AI to replace my voice. I want it to make my voice less
       afraid of trying the odd version first.
0.13 - A first draft is a negotiation between taste and momentum.
0.11 - Sometimes the best use of a model is asking it to be less sure.

I would not paste any of these directly into an essay. But I immediately know which doors are more interesting. The third one has a tension I like. The fifth one has a shape I might steal. The first one says the thing plainly enough that it could become the anchor.

That is already a better writing session.

How this post used it

For this article, the more useful prompt was not “write the article.” It was:

I want to write a short, personal blog post about verbalized sampling and using
AI to write more creatively.

Generate 8 possible frames for the piece. Each frame should feel honest,
specific, and not like a generic AI essay.

For each option, include:
- a possible title
- the central claim
- one concrete example
- a rough probability

After the list, name the 2 least obvious frames that still feel grounded.

The output I wanted was not an article. It was a set of possible articles:

0.27 - Ask for the other drafts
       AI is useful before the draft hardens. Show title, opening, and sentence
       revisions as probability-weighted options.

0.21 - Let the model stay uncertain
       The creative value is in delaying closure. Use AI to keep multiple
       versions alive long enough for taste to react.

0.16 - The hidden range of AI writing
       The model is not out of ideas; it is over-optimized for the safe one.
       Compare a default blog intro with tail samples.

0.12 - A field, not a hallway
       Ask for a distribution so writing feels less linear. Use the field as a
       metaphor for drafting with more optionality.

0.09 - Write with the tails
       Low-probability options are where the weird, useful phrasing often lives.
       Show sentence-level rewrites under 0.10.

0.07 - The almost-answer is the point
       The best model output may be the option you do not use directly, but that
       tells you what you wanted.

This post mostly took the first frame. But the lines I liked came from the lower probability options: “before the draft hardens,” “a field, not a hallway,” and “the almost-answer.” That is the pattern I want. The likely answer gives me the structure. The tail gives me the energy.

The title went through the same process:

0.30 - Ask for the other drafts
0.19 - Let the model stay uncertain
0.16 - The hidden range
0.12 - A field, not a hallway
0.08 - Write with the tails
0.06 - The almost-answer

I picked the most likely title, which is allowed. The point is not to always choose the weirdest answer. The value is that I got to see the nearby possibilities before choosing.

The tail is where it gets interesting

The repo’s quickstart also suggests asking for lower-probability responses:

Generate 5 responses. Each response should include text and a probability.
Please sample from the tails of the distribution, such that the probability of
each response is less than 0.10.

This is the part I find most useful for creative work. The first few answers from a model are often variations on the same social script. The tail prompt nudges it toward options that are still plausible, but less automatic.

For example, if I am stuck on a sentence like this:

AI helps me write by giving me more versions of a thought.

I might ask:

Generate 6 revisions of this sentence in a quiet, first-person style. Preserve
the meaning, but make each revision come from a different angle. Include rough
probabilities and include at least 3 options below 0.10.

The useful results might look like:

0.22 - AI helps by giving a thought a few more lives before I choose one.
0.18 - AI makes it easier for me to see the same idea from several distances.
0.14 - The value is not the sentence it writes, but the extra versions it lets
       me compare.
0.09 - It lets me stay undecided a little longer, before one phrasing hardens.
0.07 - It turns a single sentence into a small weather system.
0.04 - It makes language feel less like a hallway and more like a field.

Again, I do not want to outsource the choice. I want more material for taste to react to. In this set, the line I would keep working from is probably:

It lets me stay undecided a little longer, before one phrasing hardens into the only one I can see.

That feels like me. Or at least like something I would want to make more mine.

A better prompt for my own writing

The pattern I want to keep using is not “write this for me.” It is more like:

You are helping me think through a piece in my own voice.

Generate 8 possible directions for the next paragraph. Each should be concise,
specific, and honest. Do not make them sound like marketing copy.

For each option, include:
- text
- probability
- why this direction might be useful

After the list, name the 2 options that feel least obvious but still grounded.

This gives me a menu without pretending the menu is the meal. I can ask for intros, titles, examples, metaphors, counterarguments, transitions, or alternate endings. The important part is that I am asking for a range of possible moves, not one polished completion.

I especially like this for places where writing gets prematurely narrow:

• finding a first sentence
• naming what a post is really about
• generating examples that are concrete instead of generic
• trying a stranger analogy without committing to it
• rewriting a paragraph without sanding off the voice
• asking for objections before the piece becomes too pleased with itself

The caveats matter

The probabilities are not magic. The repo is careful about this: treat them as sampling weights, not calibrated truth. A model saying something has probability 0.07 does not mean the universe has spoken. It means the model is giving you a rough way to sort and sample its own possibilities.

There is also a cost. Asking for a distribution uses more tokens than asking for one answer. And if the prompt asks for too many options, quality can drop. The paper’s practical framing is refreshingly modest here: use a small k, tune the probability threshold, and remember that temperature and top-p are still separate knobs.

The biggest caveat is taste. Verbalized sampling can widen the search space, but it cannot tell me which sentence I actually believe. That still has to happen in my own head, with my own ear.

Good.

That is the part of writing I do not want to automate away.

The small shift

The sentence I am taking from this is:

Ask for a distribution, not a sample.

That feels like a useful mental model for writing with AI. Do not ask the model to collapse the field too early. Ask it to show the field. Then walk around in it for a while.

The better draft is often not the most likely one. It is the one that helps me notice what I was actually trying to say.