Genres

Bots that work alongside clinicians and care teams. They summarise, structure and cross-check — they do not diagnose and they do not prescribe. Every card in this genre is expected to carry its limits in plain language.

Bots that live in documents: drafting, summarising, keeping a house style, and answering from a body of material rather than from the open web. The ones worth having cite the page they took it from.

Bots that pull series, fit models, decompose risk and write the notebook that shows their work. Strong opinions about data provenance, no opinions about your portfolio.

Bots for legal operations: clause extraction, redlines against a house standard, obligation tracking and jurisdiction comparison. Assistance for lawyers, not advice for clients.

Bots that live in a codebase. They read the repo before they write, they open pull requests instead of pasting diffs, and they leave the test suite greener than they found it.

Bots that search a corpus, extract structured claims, chase citations back to the source and flag where the literature disagrees with itself. Optimised for being checkable.

Bots for the wet lab and the pilot line: protocol drafting, run logs, instrument export parsing, QC charting and the notebook entry nobody wants to write at 19:00.

Bots built around a cadence rather than a task: rehearsal partners, review facilitators, training-plan keepers. They ask better questions than they give answers.

Bots that keep an organisation legible: reconciliation and close, vendor and contract tracking, SOPs, incident runbooks and the operating review nobody has time to assemble.

The honest bucket. Experiments, hybrids, one-offs and bots whose genre has not been invented. When enough cards cluster here, they graduate into a genre of their own.