Your voice workflow

Your voice is where Skryba collects and measures your writing. A segmented picker chooses Samples, Harvesters, Rules, Profile, Write, Progress or Evaluate. Every run is one CLI invocation with its command and JSON result shown. Numeric evaluation controls start empty and must be supplied; CLI refusals appear in their own words.

Samples

Plan asks skryba writing-style samples plan --pairs <N> --length-bands <N> and shows each kind of writing in each length band with how many pairs it needs, how many you kept, how many wait for you and how many are still to write. Generate adds Temperature and Seed and runs samples generate, which has the Brama model write those samples; the app's environment must carry BRAMA_URL, SKRYBA_MODEL and BRAMA_TOKEN_ROLE as it does for Rewrite. List shows every sample and whether it was kept.

Show opens one sample by its id: the model's text above, and below it an editor that starts as that text. Rewrite it in your own words and press Keep mine; the app runs skryba writing-style keep with {"sample": <ID>, "kept": <your text>} on stdin, so your prose never appears on the command line, and shows the CLI's answer under the button. Remove deletes a sample nobody kept once the confirmation toggle is on; without it the CLI names the sample and removes nothing, and that refusal is what the window shows.

Harvesters

Add registers a harvester: a name, a kind (files, messages, notes, transcripts), the path it reads — empty for Transcript Lake's own data — and the fewest and most words a passage keeps. Edit changes them, and Forget where it stopped reads the source from its start next time. Remove stops harvesting; the passages stay. Run reads every harvester, or the ones named one per line, and shows what each read, added and skipped. Schedule set has Stado run every harvester on a cron in a time zone on the host you name; Schedule show reads it back from Stado and Schedule remove deletes it, then names the removed schedule and says whether Stado removed it now or had already deleted it, in which case only Skryba's record was cleared. What each kind reads and every refusal are on writing-style harvest.

Write

Next returns pending work of the chosen kind first, without touching any source; only when none is pending does it refresh all registered harvesters and create one assignment from an unused authored seed. For original writing it shows a topic, audience and purpose. For rewriting it shows the model passage. The editor sends your own response on stdin through writing-style work submit, never in the command line. List shows pending, completed and skipped work; Show opens one by id. The assignment panel can submit or set aside a pending item, and resume a skipped one. An answer typed in the panel belongs to the assignment shown: opening another assignment starts with an empty answer, and Submit my writing sends the text under the id it was typed for. The same actions are available in the form. A failed source, missing model route or exhausted authored corpus is shown as the CLI refusal rather than an empty assignment.

Progress and evaluation

Progress runs writing-style progress. It shows passages and words by kind, assignment lifecycle, the latest exact adapter digest and every held-out comparison separately. Select a run to read its persisted references, generations and verdicts through progress --evaluation <ID>, or enter an id in the form. A corpus count is never labelled a style score. The latest training row can fill the adapter path in Evaluate. Then supply the Ster device, precision, chat template, generated-token budget, lead words, held-out passage limit, temperature and seed. It runs the base and adapted local checkpoint on passages imported after training, asks Brama to compare them blind against your prose, and shows each reference, output and winner with aggregate wins and ties. A refusal for missing provenance, no later prose, unknown evaluation id, empty generation or malformed judge answer appears verbatim. The CLI contract is writing-style learning.

Rules

List, Add, Edit and Remove manage the writing rules: a name, what it measures, at most or at least, the limit, and for a phrase rule the phrase. Check counts every rule on the text you paste — it is read on stdin — and shows each rule's verdict with every sentence or paragraph that broke it. Every rule held is also counted on every Rewrite candidate. What each metric counts is writing rules.

Profile

Runs skryba writing-style profile with any quantiles you list, comma-separated, and shows the smallest, mean and largest sentence, word and paragraph length across your held passages — the values a rule is set from.