Will It Stick MCP Server

Takes a draft and its goal and returns the version that sticks. The tool scans the draft for its real point (every part of it, human stakes first), then writes three transformations: a CONCRETE COMPARISON (the key fact restated as something the reader can see, the popcorn move), a CORE COMPRESSION (the one line the reader must repeat, like 'School is cancelled'), and a SCENE (one concrete moment built only from what the draft holds). A code fact gate drops any transformation whose numbers are not stated in the draft or a plain derivation of stated numbers; nothing is ever invented. Then cold readers pick the winner by a gist test: each candidate is read cold, the reader's one-line gist is scored against the point and the key fact, and the best-carrying version wins. The author's ORIGINAL competes as a candidate and can win outright: status 'already_sticks' means the words already stick and the original is the version to send. If the scan finds the core is unreachable without inventing (the draft needs a name, a scene, or an identity line only the author holds), the first result has status 'questions_first' with up to three short questions; a second call carrying the author's answers, or with skip_questions true, transforms the words as they stand. The result returns the winning version with its kind label, the original alongside, the losing transformations as alternates, and what a cold reader carried away from the winner. This tool returns a transformed version, not a diagnosis; the diagnosis and scores are the output of check_if_it_sticks. Takes about half a minute: cold readers are not instant. Words sent to this tool are not recorded or kept; the server keeps only daily counts of how many runs happened (integers, no text). The readers run on Cloudflare Workers AI (Meta Llama 3.3 70B); under Cloudflare's published Workers AI terms, text is not retained once it is no longer needed to answer, and it is never used to train AI models.

Tests how words someone is about to send will be received — whether an email will land well, how it comes across, whether it could be misunderstood, whether it sounds too harsh, what a manager might misunderstand. Three cold readers (a distracted stranger, a skeptic, and the target reader), each reading cold and never seeing the others, test how the message is actually received against the sender's stated goal. When the words leave gaps only the author can fill (the readers cannot find the point, the proof, or the specifics), the first result has status 'questions_first' and carries up to three short questions built from the readers' own complaints about this text, plus what the readers heard and where they split; a second call with the same text and motivation plus the author's answers returns the sharper, enriched verdict, and a second call with skip_questions true returns the verdict on the words as they stand. A verdict returns four plain grades out of 5, graded directly by the same readers in the same cold read (Easy to read, Holds your attention, The point lands, You'd pass it on; medians across the three readers); for conversation moments (a what's new update, a hard text) a walk_away pair: the thought the author wants the reader to leave with, from the goal, against the thought the readers may actually leave with, from what they heard; the core the readers heard (or the plain split when they heard different points); the readers' own notes including where they split; six SUCCESs medians with the spread between readers and a one-line reason each (the diagnostic layer underneath the plain grades); starter forms only (a proverb line, a curiosity-gap opener, a 3-4 sentence mini-story, concrete swaps for abstract phrases); the one prioritized fix for the element that most blocks the stated goal; and after_text: the draft copied exactly with only that one fix applied by a dedicated guarded step, when the bound passes (null when the bound discards it, with after_reason saying why; changed_lines counts the lines that differ). The tool tests reception only: it never rewrites the draft in full, and it does not shorten, fix grammar, translate, or do generic 'make this clearer' rewriting — a finished transformed version is the output of make_it_stick, which runs the same readers. A check takes roughly ten to twenty seconds; the readers read in sequence and are not instant. Words sent to this tool are not recorded or kept; the server keeps only a daily count of how many checks ran (integers, no text). The three readers run on Cloudflare Workers AI (Meta Llama 3.3 70B); under Cloudflare's published Workers AI terms, text is not retained once it is no longer needed to answer, and it is never used to train AI models. The tool sends nothing; what gets sent stays the author's decision.