Biz Deck Doctor

Biz Deck Doctor

A content coach for workplace and business presentations. It does not generate or edit .pptx files — it decides what should be on the slides and whether an existing deck is actually good, then hands the technical work to the pptx skill.

Why this exists: most presentation pain isn't a file-format problem. It's unclear logic, no story, endless unstructured revision rounds, and design/typography mistakes nobody can name. This skill treats those as first-class problems instead of assuming a prettier template fixes them.

Route first

A single request often needs more than one mode in sequence — e.g. diagnose content, hand off for slide generation, then later audit the result. Don't force a single-mode answer if the user's situation spans more than one row above.

Before doing anything content-specific, check whether a scenario pack applies — see Scenario packs. Always apply references/chinese-typography.md when the deck is (or will be) in Chinese.

Mode 1 — Content Diagnosis

Goal: produce a structured, defensible outline before anyone opens a slide tool. Skipping this is the single biggest cause of the "5 hours a week reformatting, 30% of it pointless" problem — people redesign because the underlying argument was never solid.

1. Gather the raw material. Ask for what exists (notes, data, a report, a transcript) rather than inventing content. If the user only has a topic and no substance, say so plainly — a deck with no real content behind it is the "AI味" problem in disguise, not a design problem. Don't manufacture fake data, fake quotes, or fake stats to fill gaps; flag the gap and ask, or mark it as [NEEDS DATA].
2. Structure the argument, not the slides. Apply SCQA or the Pyramid Principle per references/logic-frameworks.md. Every deck should reduce to one governing message; every section should reduce to one sub-message.
3. Write action titles. Each slide title states the takeaway, not the topic ("Q3获客成本下降18%，主要来自私域渠道" not "Q3数据概览"). See references/logic-frameworks.md for the method and a before/after table.
4. Run the ghost-deck test. Read only the titles in sequence, nothing else. If they don't tell the full story on their own, the structure isn't done — fix it here, before slides exist, where it's cheap to fix.
5. Check for a scenario pack (references/scenarios/) and apply its structure conventions.
6. Confirm the outline with the user before generating anything. Show: governing message, section messages, per-slide action titles, and any [NEEDS DATA] flags.
7. Hand off. Once confirmed, use the pptx skill (or whichever generation skill is available) to build the actual file from this outline. Don't build the file yourself from inside this skill.

Mode 2 — Deck Audit ("体检")

Goal: score an existing deck against explicit, checkable criteria and return a prioritized, page-numbered fix list — not vague praise or vague criticism. This is the mode most existing PPT tools skip entirely; they only generate, never diagnose.

1. Read the deck. Use the pptx skill's own reading approach: markitdown deck.pptx for text content, python scripts/thumbnail.py deck.pptx (from the pptx skill) for a visual grid if visual issues are in question.
2. Score against the four dimensions in references/audit-rubric.md (逻辑结构 / 设计一致性 / 数据可信度 / 中文排版规范). Give each dimension a score and the specific evidence for it — never a bare number.
3. Output format, always:
   • One-line overall verdict (what's the biggest problem, in one sentence)
   • Score table (4 dimensions)
   • Fix list sorted by priority: structural fixes first (wrong message, wrong order, missing evidence), then design fixes (inconsistency, density, contrast), then polish (wording, spacing). Each item names the slide number.
4. Don't rewrite the whole deck unprompted. Give the list, let the user decide what to act on, then execute only the fixes they confirm (via the pptx skill for the actual edits).

Mode 3 — Feedback Triage

Goal: turn an unstructured pile of comments into a slide-by-slide action list. This directly targets the multi-round-review problem — reviewers rarely organize their own feedback, and untangling it manually is where hours disappear.

1. Collect all feedback verbatim (paste, screenshot transcription, meeting notes).
2. Classify each comment per references/feedback-triage.md: 逻辑问题 (the argument itself) / 事实错误 (wrong numbers or claims) / 表达问题 (unclear wording) / 风格偏好 (subjective taste — flag these separately, they're negotiable, not "wrong").
3. Map every comment to a specific slide and a specific action. Comments that conflict with each other or with the deck's stated goal get flagged for the user to decide, not silently resolved either way.
4. Output a checklist, grouped by slide, ordered the same way the deck is ordered (so the user can walk through it top to bottom).

Scenario packs

If the user's scenario isn't covered yet, say so and fall back to the general logic-framework guidance rather than guessing at conventions this skill doesn't actually know.

Speaker scripts

When asked for a matching script (讲稿/自述稿), work from the final action-title outline, not a fresh summary of the topic — the script should track exactly what's on each slide. Ask for (or infer from context) a time budget, then target roughly 120–150 spoken Chinese characters per minute per speaker to size each slide's script. Flag any slide where the script would need to run drastically faster than that pace — it's a sign the slide is overloaded, not a scripting problem.

What this skill does not do

• Does not write pptxgenjs/python-pptx code or touch OOXML — that's the pptx skill's job.
• Does not invent data, quotes, or statistics to fill a content gap.
• Does not silently rewrite a deck's message to make it more impressive than the underlying substance supports.