TKEX The Knowledge Exchange

Clinical Communication Course · practice tool

The Encounter

One simulated consultation. You talk, they answer, nobody coaches you while it is happening unless the level says so. Afterwards you get your own sentences back, with what each one did.

1. Pick the level

Each level turns on a different part of the encounter. The capstone turns everything on and takes the supports away.

2. Pick who you are seeing

3. Set yourself a constraint, if you want one

Optional. Nobody sees this but you. Picking a hard one and missing it teaches you more than not picking one.

4. Set up your voice and your connection

Kept in this browser only (localStorage), never sent anywhere except the Anthropic API. For a cohort, deploy the included proxy instead and leave this blank.

Length is recorded either way but never scored. Set it to your real clinic slot if you want the constraint, or turn it off if you would rather work on the skill first.

Off by default, because the stock voices sound synthetic and a bad voice is worse than none. On a Mac, System Settings, Accessibility, Spoken Content, System Voice, Manage Voices, then download an Australian or British Premium or Enhanced voice. They are free, they are a large improvement, and they appear in this list once installed.

Your voice in works three ways: the mic button below the text box, your operating system's dictation, or Wispr. They all type into the same box.

Advanced: models and endpoint
Open a saved encounter

Load a .json exported from a previous encounter and see its debrief again. Nothing is sent anywhere and no key is needed. This is also how a second human marker reads a transcript alongside the scores it was given.


Mark one blind instead. For the reliability study: you score the transcript against the anchors without seeing what the model gave it, then compare and export the pair as CSV.


Content pack. The patients, levels, anchors and questionnaire items can all be replaced with a JSON file, so content can be edited and versioned without touching the code. Whichever pack is loaded is stamped onto every record, so a change partway through a cohort is visible in the data rather than invisible.


Study export. One file with everything: consent, participant ID, tool and content versions, every encounter, and the engagement record. This is the file a study team asks for.