Riso Design v4.0

Resolve one Python interpreter and reuse it. Keep outputs outside this skill. Run python scripts/doctor.py --out OUTPUT once per environment; real renderer probe results are cached. Read harnesses only for setup failures.

Stage 0 — mandatory, shared by poster and deck

Read content planning before authoring. State purpose, audience, thesis and CTA; inventory source atoms, then group and rank them. Use verbatim for final copy, adapt for speeches/articles. Never invent facts or labels. Retain every kept/merged/condensed/cut atom in the ledger; protect numbers, names and quotes. Write content-plan.json; the command below lints, compiles and runs with source binding. Existing-spec conversion uses --from-spec INPUT.json in place of --plan content-plan.json. --skip-plan is an explicit, recorded reproduction escape; never claim Stage 0 passed.

python scripts/poster_run.py --plan content-plan.json --out OUTPUT --auto

Poster

New content plans default to mode: infographic; dense unchanged old specs use mode: editorial. Input formats describes compatibility. Read the computed source profile first; never add data to make an infographic. Budgets and text/data balance interpolate with evidence level; body ≥18pt on A3. See evidence scaling; overrides need a recorded reason. Low-data sources use source-bound typography, definitions, quotes and contrasts. Change grouping/columns before reducing copy. Adaptation requires ledger entries.

Information types choose native evidence forms: statistic, comparison, process, part-whole, timeline, quote and logistics. Data drives drawing; do not place shapes manually. Default elements are enabled: use them, do not omit them. Set only art_direction.element_density: sparse | normal | rich (default normal). See element catalogue for data forms and glyphs.

• Do bind a glyph to a claim; do not float an unrelated motif in empty space.
• Do provide ordered step data; do not draw connectors one tool call at a time.
• Do inspect quota warnings; do not cover reading text to satisfy an element count.

Use --draft for a text/geometry PNG without LibreOffice. It is not a final render. Use --variants --draft to compare columns once; --variants --auto batches conversion. Use --preflight for capacity diagnostics. --describe-spec exposes controls. Routing flags: --system or --autodiscover, or --kit PATH --node PATH. --mode audit bypasses render cache. Detailed evidence lives in files; summary.json is the entrypoint. Both runners default to the system route. See performance and cache. For editorial overrides read art direction.

Deck

Use the same Stage 0 with target: deck: one assertion and 1–4 source-bound modules per slide; use module page to group explicitly and keep narration in notes. Evidence selects page types; all new pages share the poster form writer. --draft skips LibreOffice. Existing page types remain available; read scenarios and only the relevant archetype.

python scripts/deck_run.py --plan content-plan.json --out OUTPUT --auto

Final inspection and delivery

Inspect full PNG, thumbnail and blur. Scaffold only prefills provenance, never scores:

python scripts/poster_review.py --png OUTPUT/preview/page-0001.png --out OUTPUT/visual --plan OUTPUT/poster.plan.json --review OUTPUT/visual/review.json

Fill reviewer/viewed fields, 9 scores and 0–5 located weaknesses after inspection. Keep the hash-bound spec/plan/style/QA/manifest/review chain. Missing or stale PNG is UNVERIFIED. Deliver editable PPTX, source + content plan/ledger, plan/style/media, QA, timings, summary and available PNG/PDF/review. Geometry metrics do not establish aesthetics. LibreOffice does not establish Microsoft PowerPoint rendering. No external images, downloaded fonts or network calls are needed by the pipeline. Maintenance tests live in the source checkout, outside the blank runtime package. After engine edits, run its regression suite and render representative poster/deck inputs. On demand: design system, palettes, typography, delivery checklist.