You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
scripts/check-deck.mjs: a one-command check for a finished deck. It prints one line per defect (content off the slide frame, clipped or overlapping text, text under 18px, fonts that did not load) and can write a single contact sheet of all slides. The skill now uses it for the visual check instead of ad hoc render scripts.
Improved
Lower agent token use. New-deck steps and the approval gate moved into references/new-deck.md, so edit, export and deploy runs load less. Code-slide and media guidance moved to their own files, repeated rules were removed, and maintainer-only text was cut from the export references.
Adding slides to an approved deck now uses targeted edits at a <!-- more slides --> marker instead of rewriting the file. Edits to an existing deck need no reference files unless they add a new exhibit type, code or animation.
One type scale: deck-template.md sets 80px titles, 40px subheads, 28px body and 20px labels, with floors of 26px body and 18px labels.
The root SKILL.md is generated from slide-wright/SKILL.md by tools/sync-root-skill.sh, and CI fails if it is out of date.
Limits
Measured on small samples (1-3 runs per case): about 4-19% fewer tokens per run depending on the task; the preview step changed least. With a real browser check, runs using check-deck.mjs used about 62k tokens against about 95k when agents wrote their own render scripts (2 runs each).
check-deck.mjs flags intentional overlaps such as ghosted numbers behind a title, and cannot see text blocks that touch without overlapping; the contact sheet covers those.