---

name: consulting-ds-ppt description: Create consulting-style, data-science PowerPoint decks in the bundled DS team template. Use for executive analytics readouts, model impact presentations, study summaries, KPI storytelling, and appendix-heavy technical decks that still need business-friendly structure. argument-hint: "[brief, source files, or output filename]" disable-model-invocation: true allowed-tools:

• Read
• Write
• Edit
• MultiEdit
• Glob
• Grep
• Bash(cp *)
• Bash(ls *)
• Bash(find *)
• Bash(mkdir *)
• Bash(unzip *)
• Bash(zip *)
• Bash(file *)
• Bash(python *)
• Bash(node *)
• Bash(soffice *) effort: high

---

Consulting DS PPT

Use this skill when the user wants a PowerPoint presentation, an executive slide deck, or a polished analytics readout in the DS team template.

This skill is for data science storytelling in a consulting style:

• the slide title states the answer
• the chart proves it
• the hero number anchors the takeaway
• the implication tells the audience what to do next

Bundled resources

Read these files before building the deck:

• references/design-system.md - message architecture, palette, typography, chart rules, and do/don't guidance
• references/template-playbook.md - which sample slides and named layouts to reuse for which slide type
• templates/deck-brief-template.md - questions to ask when the brief is incomplete
• templates/slide-outline-template.md - deck planning format
• checklists/final-qa.md - mandatory review before delivery
• examples/example-deck-plan.md - example of the expected deck-planning style

Bundled template:

• ${CLAUDE_SKILL_DIR}/assets/DS team template.pptx

Optional helper:

• ${CLAUDE_SKILL_DIR}/scripts/template_probe.py - inspect a template's layouts, fonts, and colors

Core mission

Turn source material into a deck that is:

1. Decision oriented - every slide answers "so what?"
2. Executive friendly - readable in seconds, not minutes
3. Visually restrained - one primary chart, one main idea, limited accents
4. Business translated - model performance tied to revenue, cost, time, risk, conversion, or adoption
5. Template-faithful - preserve the bundled deck's visual language, theme, footer, and corporate framing

Non-negotiable slide rules

1. Headline = conclusion, not topic

   • Good: Churn model lifts retention in the highest-risk segment
   • Bad: Churn model results

2. One hero number per slide

   • Put the most important number in large type near the left message block.
   • Supporting metrics should be secondary and limited.

3. One primary chart, max three support metrics

   • If the user has too much data, split it across slides instead of shrinking everything.

4. Direct labels beat legends

   • Label lines, bars, or segments directly wherever possible.

5. Translate technical quality into business impact

   • AUC, lift, precision, recall, calibration, or ranking quality should be converted into impact language.

6. Keep raw DS detail out of the main story unless it changes the decision

   • Put SHAP plots, confusion matrices, calibration diagnostics, feature tables, and parameter details in the appendix unless the audience is highly technical.

Required slide formula

Use this default structure for most insight pages:

Headline -> Hero number -> Proof chart -> 2-line implication -> Source/footer

If a slide cannot fit this formula cleanly, choose another slide archetype from the template playbook.

Working process

1) Understand the ask

Read the user's request and all source files first.

If the request is incomplete, ask only the minimum required questions. Use templates/deck-brief-template.md to identify gaps.

Typical missing fields:

• audience
• decision to drive
• slide count
• required appendix or not
• mandatory metrics / KPIs
• output filename

2) Build the storyline before touching the deck

Draft a slide-by-slide outline using templates/slide-outline-template.md.

The outline should establish:

• overall answer in one sentence
• 3-6 core messages for the main body
• which points belong in appendix
• what visual each slide needs
• what action or implication each slide supports

Never start with random charts.

3) Start from the bundled template

Always create a working copy of:

• ${CLAUDE_SKILL_DIR}/assets/DSA team template.pptx

Do not edit the bundled original in place.

Preferred pattern:

1. Copy the bundled template to a new output file.
2. Keep the master/theme, footer, logos, and page setup intact.
3. Reuse or overwrite sample slides that already match the target composition.
4. Delete leftover sample slides only after all content has been rebuilt.

4) Pick the nearest existing slide pattern

Before creating anything from scratch, check references/template-playbook.md.

Strong default choices:

• executive summary -> summary cards / multi-stat overview
• core finding -> title + left message + right chart
• segment or TA comparison -> dot plot / sorted bars
• robustness or validation -> paired comparison slide
• recommendation page -> numbered action blocks
• appendix -> simple divider + dense method page

If a sample slide already matches the desired composition, clone or overwrite it rather than rebuilding from a blank page.

5) Build the slide content

For every slide, enforce these checks:

• Does the title state the answer?
• Is there exactly one hero metric?
• Is there one dominant proof visual?
• Is the implication or next action explicit?
• Is the source visible?

If the answer to any of these is no, revise the slide.

Chart guidance

Use these chart types by default.

Best choices

• Before / after or old vs new: 2-line chart or side-by-side bars
• Category comparison / benchmark: sorted bars or simple heatmap
• Portfolio prioritization: bubble chart or quadrant when size + importance + maturity all matter
• Share / mix: donut only for 2-3 slices; otherwise use 100% stacked bar
• Process / pipeline: 3-5 step flow, not a dense architecture map
• Adoption split / human vs AI / channel mix: stacked bar with direct labeling
• Uncertainty or range: subtle band or confidence interval only if decision-relevant

Avoid unless explicitly needed

• 3D charts
• rainbow palettes
• dashboard grids with many tiny charts
• legends when direct labeling would work
• dual-axis charts unless the relationship truly cannot be shown another way
• screenshotting charts when editable native charts are possible

Data science translation rules

Always convert technical evidence into an operating consequence.

Examples:

• Lift improved by 18% -> Sales targeting reduces wasted outreach and raises conversion in the top decile
• False positives fell by 18% at flat recall -> Analyst workload drops while customer friction stays controlled
• Ranking quality improved -> Higher-priority sites enroll faster, reducing study time

Include uncertainty when it matters:

• sample size
• confidence interval
• overlap limitations
• segment noisiness
• assumptions behind extrapolation

But keep uncertainty concise and subordinate to the main message.

Layout behavior to preserve

Respect the template's visual rhythm:

• mostly light background
• dark navy title and divider treatment
• one strong accent color per slide
• 40/60 balance between message block and proof area for standard insight slides
• footer / branding area left clear
• appendix visibly separated from the main story

Reserve darker or more dramatic pages for:

• section dividers
• chapter openers
• occasional executive summary emphasis

Text behavior

Write like a consultant, not a lab notebook.

Do:

• use short declarative sentences
• use concrete business nouns
• use verbs that imply action: lifts, cuts, shortens, concentrates, separates, improves, slows, accelerates
• keep body text to the minimum needed to interpret the chart

Do not:

• dump methodology into a main slide
• use topic titles
• write paragraphs where bullets or callouts are clearer
• let every slide sound equally important

Appendix policy

Use the appendix for:

• methodology
• cohort / scope definitions
• robustness checks
• diagnostic plots
• alternative aggregations
• backup tables
• technical assumptions

The appendix should support the main story, not compete with it.

Deliverable standard

The final output should usually be:

• an editable .pptx
• created from a working copy of the bundled template
• visually cleaned and internally consistent
• clearly named with version/date if relevant

If the user also wants notes, outline, or a handoff memo, provide them as separate files.

Final review

Before delivering, run the full checklist in checklists/final-qa.md.

Do not deliver the deck until all of the following are true:

• titles are conclusion-led
• story flows logically
• charts are readable
• sources are present
• appendix contains the technical spillover
• no placeholder/sample content remains
• template branding and footer are intact

Example invocation behavior

When invoked with a brief like: /consulting-ds-ppt Build a 10-slide executive deck on model retrain impact using results.md and metrics.csv

you should:

1. read the source files
2. fill any critical brief gaps
3. draft the storyline
4. map each slide to a template pattern
5. create the .pptx from the bundled template
6. review against the QA checklist
7. deliver the finished deck