AutoDesign PPT

Create the deck with the host coding agent. Use the bundled harness for source evidence, durable attempts, deterministic QA, editable PowerPoint export, and truthful delivery. This Skill does not require AutoDesign, a server, or an API judge.

Read first

Read source-grounding.md, output-contract.md, and review-rubric.md completely. Resolve SKILL_ROOT to this installed Skill folder. Put the run in a user-selected output directory outside the Skill. Never write output, caches, or run state into the installed package.

Execute

Resolve a Python 3 launcher first: try python3, then python; on Windows also try py -3. Use the first command whose --version reports Python 3 in place of <PYTHON3> below. Run <PYTHON3> "$SKILL_ROOT/scripts/ppt_harness.py" --help for exact flags.

1. Run doctor, then init --run-dir "$RUN" --source "$PAPER". The first online run installs artifact-hash-locked browser and native-PPT dependencies in a versioned, content-verified user cache; later runs can reuse them offline. A missing or changed runtime is blocked, not a reason to skip QA.
2. Use evidence --query "...", visuals, and rendered paper pages. Query evidence for identity, problem, gap, contributions, method, results, and conclusions. Query ablations, robustness, qualitative evidence, and limitations only when the paper actually reports them. A fresh host VLM or subagent must inspect PDF visual candidates against their captions before using them; bind the complete hash-matched JSON with bind-visuals. Uncertain candidates stay unused.
3. After the evidence queries, write a complete role/evidence story-plan JSON and run plan --brief "..." --story-plan "$STORY_PLAN". It has exactly one entry per slide with slide_id, a role valid for that narrative slot, and one or more real evidence_refs; the harness validates role-to-evidence compatibility before the immutable hash is written. The problem -> method -> evidence -> takeaway backbone stays ordered. Conditional experimental slots without source support must use an allowed same-phase source-backed role such as method detail, results deep dive, evidence analysis, case analysis, or scope and boundaries; never invent a robustness, ablation, qualitative, or limitations section just to fill the arc. If the story plan is omitted, deterministic semantic role scoring selects evidence and performs the same evidence-conditioned substitutions; it never assigns by extraction order. Ablation requires explicit ablation terminology, or a component/variant comparison with a measured effect; bare words such as without or variant do not qualify. Robustness and qualitative roles likewise require explicit terminology or a condition/case plus an observed result. A two-value numeric comparison can provide the observed result when the component operation or shifted condition is explicit. Reject negated, not-provided, or future-work statements before evaluating positive signals: stating that an experiment is absent is not evidence that it was performed, including localized declarations such as 本文缺乏消融实验 or 本文无定性分析; do not confuse those with method terms such as 无监督 or a measured result from removing a component. Evaluate each independent sentence or clause on its own: a complete non-negated positive clause is sufficient, but never assemble an operation from one clause and a result from another. The explicit w/o/full labeled numeric pair is the only supported two-clause comparison. Both sides of that pair must contain a value and share a measured metric (or an explicit percentage scale). When a metric is named, its value must be locally bound to it. Token-attached architecture values such as 2 heads, 2-stage, depth is exactly 4, head-count of 4, or head count has a value of 4 remain structure counts and do not become outcomes merely because accuracy appears elsewhere. Conversely, a distant structure phrase cannot steal a value explicitly bound to a metric, as in accuracy for the 2-stage configuration is 79.4%, or in accuracy is 79.4% across two stages. A fully unlabeled two-percentage pair may infer the common outcome only when neither side contains token-bound structure counts. Architecture counts are not outcomes. That fallback needs role-distinctive evidence with a unique winning margin; ambiguous or generic overlap stops and requests --story-plan. Paper decks default to exactly 18 slides. A count attached to the requested deck/slides/presentation, including Chinese numerals through sixty, overrides the default; source metadata such as “12-page manuscript”, “25-page article”, or “30-page PDF” does not. Build one research argument, not disconnected summary cards. Pass an explicit visual-allocation JSON to plan; keep each visual within its permitted role and reuse limit.
4. Run begin. Author the returned artifact/deck.html yourself. The HTML is canonical and the immutable plan is snapshotted under artifact provenance. Match every slide's ordered role, section, assertion, and evidence refs to that plan: the visible h1 must equal the planned assertion, every visible text/table source list must equal the planned evidence refs, and the speaker note must equal the planned note intent. Stage each planned figure with stage-visual. Use the exact tagged-DOM contract so native text, tables, images, shapes, and speaker notes can become editable PowerPoint objects. Use only local, regular, non-linked source-grounded paper assets; never invent paper facts, links, logos, figures, or measurements. Bind every visible text box, native table, and speaker-note statement to real evidence IDs.
5. Run validate. Static gates run before rendering. The harness then audits every slide in a network-denied browser, verifies that the authored slide roots actually compute to 1920x1080 before any audit isolation CSS, makes a contact sheet, verifies an exact-page-count PDF, exports deck.pptx, reopens its required native text/table/image counts, native rect/ellipse/line counts and types, and exact notes, and, when LibreOffice is available, renders and compares the PowerPoint with the canonical HTML. A screenshot-only deck does not pass.
6. Run review-context. Give the contact sheet and every individual preview to a fresh host VLM or fresh subagent that did not author the attempt. The reviewer must inspect every slide, cite localized repairs, and return the exact hash-bound schema in review-rubric.md. Record it with record-review; never self-certify from source code.
7. If review fails, run begin again and repair only the reported slides. The harness allows at most three attempts. If it passes, run finalize. Deliver only the allowlisted, verified final/deck.html, final/deck.pptx, final/deck.pdf, final/notes.json, local assets, and provenance. Extra, symlinked, hardlinked, or unreviewed files are never delivered.

Run resume --run-dir "$RUN" before every continuation after interruption. It verifies the installed skill_root, snapshotted instructions, source, artifact, preview, review hashes, the minimum score for persisted passing reviews, and single-link final files. A scaffold, fallback, stale review, failed export, or visually unreviewed deck is not a completed deliverable.

Design standard

Use a formal academic light theme with a restrained palette, serif display hierarchy, sans-serif labels, generous but purposeful whitespace, and native evidence. Keep every primary slide canvas opaque pure white (#FFFFFF) in screen and print so white-background paper figures integrate cleanly; never use a slide-root gradient or image, and do not apply filters, masks, or blend modes to a slide or its artifact ancestors. Restrained light fills and effects may distinguish local panels and cards. Prefer diagrams, plots, tables, equations, and paper figures over decorative stock imagery. Reject dark dashboards, repeated card grids, gradients, sparse slogan slides, oversized titles, tiny captions, and generic AI-marketing language. Keep references style-only; never copy their content.