Make the thinking legible.
Bad writing does not stay local in a vibecoding workflow. Vague instructions become slop, slop burns tokens, and each handoff can compound the ambiguity.
This is a lightweight heuristic checkpoint, not a model verdict. It measures visible text signals so the author can tighten the handoff.
One clean handoff beats five fuzzy ones.
Use the checkpoint before an agent prompt leaves your context. The goal is not ornate prose; it is a specific action, object, and finish line.
Is it clear?
Visible string heuristics count flags without a model judgment
Read the explanation
The source has six filler entries: thing, stuff, better, important, somehow and etc. It checks substring presence for each entry, not word occurrences. Write thing thing is three whitespace tokens but only one filler hit. Three equal word boxes count three tokens and one highlighted box counts the unique filler entry. These flags are transparent string heuristics, not model judgment or proof of clear thinking. The source adds one length flag when whitespace token count exceeds sixty. Common eight pixels per token bars compare sixty at four hundred eighty pixels and sixty one at four hundred eighty eight. If the text has an action substring and no filler, sixty yields zero flags and sixty one yields one. Sentence count separately splits punctuation. None of these rules measures meaning, factual accuracy or whether instructions are sufficient. Total flags add unique filler matches, one when no action regex matches, and one when count is over sixty. With two matched fillers, a missing action and excessive length, the count is two plus one plus one equals four. Four equal blocks show those components. Download reanalyzes then exports original typed text and the counts; saved text is trimmed. Action and filler tests use substrings, so build inside builder can count as an action. No agent or model executes the request.