1000ish

The experiment asks whether a model can describe an image in exactly 1,000 words. A prompt alone was not enough, so I built a loop that counts the draft, keeps the good part, and rewrites only the ending until the final text passes validation.

The model starts with a loose target of about 1,050 words. I do not ask it to perform exact-count arithmetic. Then I keep the opening intact, cut at a sentence boundary between 800 and 900 words, and ask for a replacement ending with the remaining budget. If the first draft lands on exactly 1,000 words and ends cleanly, it ships unchanged.

A validator counts each proposed ending and tells the model how many words to add or remove. It gets three attempts. Both the server and browser check the assembled text before the interface calls it complete. If the retries fail, I leave the draft visible rather than pad it or cut it mid-sentence. The count uses whitespace-separated words, so contractions and hyphenated phrases count as one word.

The correction needed to be visible. The interface dims the original ending, backspaces through it, and reveals the replacement. I buffer and validate that replacement before animating it, so you are watching a staged revision rather than live model edits. The “-ish” disappears only after the final count passes.