pptx-template-quality
Fill an existing PPTX template with evidence-backed content while preserving its layout and reviewing rendered screenshots when vision is available.
LazyAGI · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add LazyAGI/LazyMind --skill pptx-template-qualityAbout this skill
name: pptx-template-quality description: Fill an existing PPTX template with evidence-backed content while preserving its layout and reviewing rendered screenshots when vision is available.
Template filling quality
Read every page with pptx_analyzer. Its quality_contract is this skill and its layout_checks are rough hints, never a publication gate or proof of overflow. Delivery takes precedence over visual quality. The fill tool publishes a usable PPTX first, then reviews screenshots and attempts at most two geometry-only repairs. Missing rendering/vision, review errors and unsuccessful repairs keep the last delivered version and do not fail the step. Do not rewind upstream planning or retry generation to satisfy a quality gate. Disclose unresolved issues accurately.
In generate_ppt, consume the saved requirement_analysis, material_summary, deck_outline and slide_outline. Do not search the web, knowledge base, PRs or images again. If evidence is insufficient, report the missing material and return to collection only with the user's agreement. Template analysis is local reading, not a new research step. Cache hits still require reading the returned content.
-
Use ppt_fill_template pages to plan output order and count. Repeat source_page when content needs more room; honor an explicit user page count. Keep each reused page's layout, coordinates, typography, backgrounds and branding. Remove unused placeholders, sample charts/pictures/text boxes with explicit remove_shape_ids from inspection; never infer that all artwork is disposable. Resolve content against each box's effective layout, including inherited placeholder geometry and fonts. Judge the rendered result rather than guessed capacity.
-
Allocate content by whole text box before assigning text slots. A short styled run is not a container for an entire paragraph. Preserve bullet paragraphs and paragraph levels; avoid manual bullet characters and repeated blank lines.
-
Use pages.tables for aligned lists, work records and comparisons. Supply an optional headers array, rectangular rows of plain-text cells and a box in inches. Never simulate a table with pipes, spacing or separate cell text boxes. Keep baseline/result values in their named columns; recovery never guesses or moves facts between columns. Tables have their own capacity calculation and do not consume the text-box count limit. Font reduction requires an explicit min_font_size. Correct TABLE_COLUMNS_MISMATCH or TABLE_LAYOUT_OVERFLOW by fixing the reported table, allocating space or splitting within the requested page count, not by dropping cells or falling back to unaligned text.
-
Prefer concise sentences and a few supported contributions over a PR inventory. Keep qualifiers, units and source attribution: an isolated 84% outline timing reduction is not an 84% end-to-end improvement. Separate merged and open work.
-
Replace example identity/date values only with supported facts. If a required date is unknown, use a short explicit “日期待确认”, not an invented date or XX. Do not leave a required summary empty. Do not cram metadata into table headings.
-
Keep essential citations compact. If notes editing is unavailable, retain detailed evidence in material_summary; do not pretend notes were updated.
-
Before ppt_fill_template, compare all resulting text against layout_checks. Shorten over-budget content while retaining its meaning. Do not arbitrarily shrink fonts or remove required evidence to pass. Reposition or widen body boxes explicitly using shape_boxes when the slide has free space; keep brand artwork/title/footer intact. For balanced multi-column layouts, remove obsolete body shapes and use editable text_boxes with explicit bounds. Split supported content across repeated source pages when needed.
-
The tool rejects clearly over-capacity changed boxes before writing a file. Use its page/shape diagnostics to repair only affected text and resubmit the complete action list against the ORIGINAL template hash. At most two repairs; if still blocked, report the exact limitation instead of bypassing validation.
-
Whole-paragraph replacements use newline-separated independent paragraphs; short rich-text run replacements retain inline breaks. Do not paste several unrelated sections into a subtitle. The tool resolves inherited typography.
-
Geometry/text estimates are not visual validation. Validation belongs inside generation, before publication. Successful publication completes generation: neither the subagent nor parent chat should launch post-generation inspection. Report any returned limitations briefly without reopening the workflow or claiming native visual verification. A disclosure is not a pending task.
-
For composition, use ORIGINAL template slot IDs in each page's edits, including repeated pages. Omit slots belonging to removed shapes. Review every remaining sample value; do not keep fictitious achievements. Supply fresh notes per page; composition clears sample notes and internal slide navigation.
-
Use editable column/bar/line charts for supported performance comparisons; use pie/doughnut only for mutually exclusive parts of one known whole, with one nonnegative series and a positive total. Never turn unrelated improvements into pie slices or invent time-series observations. A key quantitative claim with complete evidence should have a chart, not only a large KPI number. Supply categories, named numeric series, unit, title, evidence source and box in inches. Delete the sample chart/body placeholder occupying the target area first. Preserve measurement scope, baseline and conditions in notes; never invent measurements or mix percentages with milliseconds. Chart citations are retained in notes. Use an available layout with sufficient space.
-
Technical/research slides must preserve substantive on-slide explanations of root cause, concrete mechanism/steps, results a