no-misreadlisted

# No Misread Two parties are talking through a keyhole and neither can ask the other what they meant. You write a prompt and the model reads something you did not say. The model writes an answer and you skim past the one sentence that mattered. Same failure, opposite directions. ## It works out the direction on its own Never ask the user which mode they want. They came here with a piece of text and a problem, not with a taxonomy. The text already says which way it is going, and `--check` reads that and reports it: ```json "direction": "going out to a machine", "direction_why": "detected: 12 of 26 sentences open with an imperative" ``` **Going out.** Prompts, instructions, tool and function descriptions, error messages, agent-to-agent text. Text a machine parses mid-run, where the reader guesses instead of asking. Ambiguity is the enemy, so even, flat, literal sentences are correct. **Coming back.** The answer, the draft, the README, the post going out under a name. Machine cadence is the enemy, so the sentences have to move like a person wrote them. Most rules below hold in both. Exactly three flip: | | Going out | Coming back | |---|---|---| | Sentence length | Even, under 25 words | Varied, or it reads generated | | Contractions | Avoid, a model misreads `won't` | Keep, people use them | | `may` `might` `could` | Avoid, the reader guesses which | Keep, confidence is content | This is not a compromise between two philosophies. One text has one job. Read `direction_