120 cases have prompts under 200 characters, and 70% of those are verbatim from the original post

By Alpha Lay · 2026-09-26

The library contains a batch of unusually short prompts. We were about to write them off as noise, then checked where they came from — and found they reveal publisher behaviour, not a collection problem.

Where the data comes from

pc is the character count of a case's prompt text. Of 1,274 records, 120 (9.4%) come in under 200 characters, with a median of 97.

For scale: the library's median prompt is 1,563 characters. This batch is a sixteenth of that.

Finding 1: 57.5% are verbatim from the original post

Provenance of those 120:

Nearly six in ten short prompts are short because that is how they were written.

This matters — it rules out one possibility: it is not our scraping truncating content. The publisher simply did not write a detailed prompt in the post.

Finding 2: Hosting is concentrated too, pointing to one batch

69 (57.5%) sit on the same Railway project, matching the creator-verbatim count exactly. Both signals point the same way: these mostly come from one collection batch, one that copied posts verbatim.

Finding 3: Skewed toward performance genres, with shorter clips

Their median duration is 13.7s, below the library's 15.2s (the prompt-length piece has the full band table).

The reasonable reading: short prompts tend to accompany single-action, single-subject content — "someone working out," "a dance" — expressible in a few sentences and with no need to fill the default length. Content that needs long prompts (action & physics, sci-fi) is inherently multi-step.

Finding 4: A data-hygiene problem we have to own

The minimum pc value is 4 characters.

Four characters cannot be a meaningful prompt. It is far more likely a field-splitting error (some identifier string mistaken for a prompt) or a very short source line treated as one.

We did not remove it, because no consistent rule for removal exists (there is no objective threshold for "invalid"). But readers should know:

What you can do with this

If you learn prompt craft from public cases: when a prompt looks suspiciously short, check whether it is "verbatim from the original post." Cases where the post only carried a title genuinely exist — they are not models to copy.

If you build a case index: this is worth recording — a missing field and a short field are two different diagnoses. The first is your problem; the second may be upstream. Always pull provenance before deciding whether to clean; we nearly deleted 120 informative records as noise.

If you publish your own cases: writing prompts out in full is in your own interest. By the data, cases with no prompt in the post can only exist downstream as a "title," and title-based classification is very weak (see the annotation-basis piece).

Method and limitations

• The 200-character threshold is arbitrary: a different cut (300, or 150) changes the count, though the source concentration holds.
• pc and title are different fields: this piece does not treat "short title" as "short prompt." The case names in the table above are titles, shown for illustration only and not the contents of the pc field.
• 57.5% from a single source weakens any generalisation to "the community does not write prompts." We claim only that such records exist and are verbatim.
• We cannot distinguish "the publisher wrote nothing" from "the publisher wrote something but did not call it a prompt." The latter gets logged as a short prompt.
• All figures use the 1,274 index records.

Reproduce it

Case detail views show the prompt and its length: https://cases.aishifu.shop/