# Tell Me When

> Turns written instructions into a talk-through for the phone. Paste the steps you have to read out
> to someone who is doing the job on the other end of a call - tying a bow tie, setting up a board
> game, a recipe, a new vacuum bag. A free in-browser engine splits the text into written steps, times
> each one at a slow, steady or brisk speaking pace, marks what trips the ear (a picture the listener
> cannot see, numbers and codes, symbols, abbreviations, asides in brackets, long sentences, several
> actions at once) and draws the steps as knots and tangles on a tin-can telephone, with a verdict by
> a fixed rule: send a picture first, two calls, slow it down or read it out. A metered run writes the
> talk-through. Derived from @jwynia/speech-adaptation.

https://tell-me-when.skillsafe.ai/

## What it does

- Written steps: one a line (numbers and bullets are dropped); a single paragraph is split into sentences. Two to twenty steps.
- Trips the ear: pointing at a picture ("as shown", "figure 2", "the diagram"); numbers and codes (three digits or more, or digits stuck to letters or symbols); symbols; abbreviations; an aside in brackets; a sentence over 22 words; two or more of "and", "then", "after that" or a semicolon.
- The verdict, checked in this order: a step points at a picture is send a picture first; else 13 steps or more is two calls (the first call ends after the middle step); else a third or more of the steps tripping the ear is slow it down; else read it out.
- A call reader shows one part at a time in large type, with an elapsed clock, for reading during the call.
- Free downloads: the drawing (SVG, PNG), the steps with their marks (CSV), and the talk-through (plain text to read from, Markdown, JSON).
- Saved talk-throughs (opt-in) keep the text, the listener, the pace and the latest talk-through, and show how the verdict moved between runs.

## The paid task

One task, `talk`, on gpt-terra, metered per run. The input is four string fields: `task`, `job` (the job, the listener and the pace), `steps` (one written step a line with its word count and what trips the ear) and `facts` (the page's numbers and choices). The reply is one JSON object: verdict (send-a-picture | two-calls | slow-it-down | read-it-out, always the page's), before_you_call, have_ready, opening, parts (each with steps, say and check), pause_after, pause_line, closing and tip. The page re-checks the reply: every written step must be in a part, in order; every number in a step must be said in its part and no number may be added; every check-in must be a question; nothing may point at a page; a break must sit after the right part; and when the steps mention heat, blades, gas, electricity, medicine or ladders, the opening must ask the listener to stop if anything seems wrong.

A reading aid, not professional advice. For anything with mains electricity, gas, medicine, sharp blades, heat or ladders, follow the maker's own instructions.

- API tutorial: https://tell-me-when.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@jwynia/speech-adaptation (jwynia/agent-skills)