Skip to content

slide-maker 6.2.0 — design strength restored and gated, a SKILL.md agents read to the end, a run the user can see

Latest

Choose a tag to compare

@dong845 dong845 released this 08 Nov 15:29
· 8 commits to main since this release

Design strength restored and gated, a SKILL.md an agent can read to the end, and a run the user can see.
A blind comparison against the older v4.2.0 showed the newer skill had become timid: every slide on one quiet
ground, no page that shouts, and agents skipping half of a 325 KB SKILL.md. This release measures that, fixes what
was measured, and holds the fixes with gates on the built deck. On the two briefs re-run with the final version
(a product launch and a city guide), blind side-by-side judging preferred it over both v4.2.0 and 6.1.0 on all eight
sheets, each pair shown in both orders. Two briefs, one run each: a direction, not a guarantee.

  • Light/dark pacing is planned and built. The Design plan's Deck rhythm ends with grounds: flips <slides> — why (two or more interior pages on the flipped ground) or grounds: single — <conservative | visual-language | template-locked | tiny-deck>: why; --gate-check reads the rendered pages and refuses a planned flip that is not
    there. Register kits paint the quiet ground on every content page, so a flip is painted on purpose with
    dk.slide_background.
  • One hero page, measured. hero: slide <n> names the page the deck exists for; its largest type must reach
    2.5x the deck's typical top size on the built file (or hero: none — <conservative | template-locked | tiny-deck | image-led>).
  • The art director's Design plan is a file the gate re-reads (scripts/design_plan_record.py): it must carry
    agents/slide-design.md's Output sections and its self-verify checks, and records whether a dispatched art
    director wrote it. On a host that can dispatch subagents, the art director and the critic are subagents.
  • Balanced density (图文均衡) by default, also for a delegated pick; a density waiver now counts only when the
    user chose text-heavy.
  • A visual language is a vocabulary, not the deck's layouts: stock page functions for the standard pages, the
    signature, densest and key-data pages composed for the content.
  • SKILL.md 325 KB → 250 KB, nothing lost. Script- and gate-backed detail and the long incident accounts moved
    verbatim into references/ (setup and gate rules into troubleshooting-faq.md, the research budget into
    content-plan-spec.md) with a named-trigger pointer left in place; the component catalogue, the render self-check
    and the design method stay inline. check_skill_lossless.py: every baseline line accounted for.
  • The user sees the run. The first full render opens the contact sheet on a desktop session (not CI, not SSH;
    SLIDE_MAKER_NO_OPEN=1 turns it off), and the commands that mark each stage print one line to tell the user.
  • Numbers. source_check.py reads the Codex evidence record, accepts a derived figure in the slide's own unit
    (万元) and treats month labels as dates; --write-rows records the unambiguous rows.
  • Brand step from a website (#8): offers easybrandkit@0.4.0 (pinned) and runs it only on a yes.
  • For maintainers: scripts/design_benchmark.py (advisory design numbers) and scripts/design_bench/
    (gallery content as test material, blind side-by-side sheets). README: the twelve example decks are back.