Mckinsey Market Research Deck

McKinsey-Style Market-Research Deck

Build a research-backed, visually elite, page-turning deck (HTML reviewed on screen → PDF for sharing). This skill is the distilled, reusable playbook. Read the four reference files as you reach each phase — do not try to hold all of it in head at once.

• references/methodology.md — how to do the research and what each section must contain
• references/design-system.md — the locked visual contract (tokens, page types, layout laws)
• ../mckinsey-deck/assets/deck_engine.py — the canonical engine (owned by the mckinsey-deck style skill; this skill consumes it — never fork a local copy, that's how drift starts)
• references/qc-checklist.md — the self-verify pass before delivery
• references/image-handoff.md — the template that hands product/cover images to an image generator

The 7-page spine (always)

0. The Answer — one answer_slide() right after the cover: the governing thought (the full recommendation in one sentence) + 3–4 pillar conclusions with key numbers. Pyramid Principle: the answer comes first; the rest of the deck is its proof. Drafted in Phase 1.5, finalized last.
1. Market Overview — size, growth, channel, the structural shift
2. Brand Landscape — Good/Better/Best ladder + brand-by-brand profiles
3. Product Categories — per-subcategory competitor price ladder + pain points + the brand's lineup
4. Customer Pain Points — sourced failure modes, each one a selling-point opening
5. Opportunities — pain points → product direction
6. The Solution — positioning, pricing/packaging, the line plan, and the decision pages (bottom-up market sizing, economics, business case) + the thesis

End with a full source register (every URL, numbered).

Workflow (run in order)

Phase 0 — Scope + Day-1 hypothesis

Get: the brand, the parent retailer/company, the category, the geography, the SKU-count target, and the strategic question (usually "what line should we build and why"). Confirm the deck is the deliverable (pure market research), not a precursor needing first-party data. Then write the Day-1 hypothesis — a one-paragraph draft of the answer ("we believe X because A/B/C") before researching. It steers the research (80/20: go deep only on the branches that confirm or kill it) and it is there to be falsified, not defended — revise it whenever the evidence disagrees, and say so in the deck.

Phase 1 — Research → one data file

Do the research per references/methodology.md. Land everything in a single <brand>-data.json (the deck is data-driven from it). Every number must carry a sourceUrl. Schema in methodology.md. Use WebSearch/WebFetch; capture competitor prices/plan tiers live with the capture date (shelf price for goods, plan/ACV for software, cost-to-adopt for OSS/service). Source bar (full rules in methodology.md § The source bar): prices from the vendor's own page only; market sizes from named research, never an SEO aggregator alone; pains quoted verbatim from a named venue; load-bearing inputs need 2 sources or an explicit "judgment call" label; floor of ≥1.5 unique URLs per content page with ≥50% primary/named-research — and zero padding URLs.

Phase 1.5 — Ghost deck (dot-dash storyline)

Before rendering a single page, write the headline-only outline: every page as one action-title sentence, in order, plus a one-line sketch of its exhibit. Then run the horizontal-logic test: read the headlines top to bottom — they must read as one persuasive essay (SCQA arc: situation → complication → question → answer). If a headline doesn't advance the argument, the page gets cut or merged now, before any layout work is spent. Draft the §0 governing thought + pillars here too.

Phase 2 — Generate the deck

Copy the canonical engine ~/.claude/skills/mckinsey-deck/assets/deck_engine.py into the project (single source of truth — engine fixes go back to that file, never to a project-local fork). Point it at <brand>-data.json, set BRAND, compose the 6-section build() (the engine ships the renderers + an example build). Render:

bashCopy code
python3 deck_engine.py                       # writes <Brand>-Deck.html
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --headless --disable-gpu \
  --no-pdf-header-footer --print-to-pdf="<Brand>-Deck.pdf" "<Brand>-Deck.html"

After EVERY build, assert structure: div diff must be 0 (an unclosed div breaks pagination).

bashCopy code
python3 -c "h=open('<Brand>-Deck.html').read();print('div diff:',h.count('<div')-h.count('</div>'))"

Phase 3 — Decision pages (adversarial verify)

The three pages that turn "opportunity scan" into "decision deck": bottom-up market sizing (TAM/SAM/SOM), economics (validate the value/margin claim with the buildup that fits the category — landed COGS for goods, CAC/payback for SaaS, adoption→conversion for OSS), business case (investment, 3 scenarios, payback). The questions are universal; the arithmetic forks by archetype — see methodology.md (§ Decision pages) for each pattern, and set exhibit.boldKeys to mark the answer rows. Generate + verify them with a Workflow pipeline — one analyst agent per page, then an adversarial verifier that re-derives every number. Persist each verified page as _decision_<id>.json.

Phase 4 — Images

List what's missing (cover hero + any product cutouts + optional dividers). Write the handoff with references/image-handoff.md, hand it to the image generator, then wire the returned PNGs into the engine.

Phase 5 — QC (mandatory, self-run)

Run references/qc-checklist.md end to end: render every page, eyeball for overlap / mid-word truncation / >30% whitespace / unblended images, and programmatically verify the source register count. Fix defects, re-render, re-check. Only then present.

Hard design laws (baked in — never violate)

These are user-confirmed preferences; treat as non-negotiable defaults:

• No eyebrow/kicker labels. No small letter-spaced ALL-CAPS tags above titles or on the cover. The action-title headline carries the meaning.
• Cover is ultra-minimal: title + one italic subtitle, on the navy hero image. No KPI band, no method/evidence block, no "Prepared ".
• Action-title headlines: the headline is a full-sentence conclusion, not a topic label.
• Charts ride beside their analysis. Embed the donut/bar/heatmap in a content page's right rail (essay_slide(side=...)); never put a chart alone on a standalone page.
• ≥4 charts per deck, each centered-titled and load-bearing. A deck has at least four charts (donut / bar / heatmap mix) that each carry a real finding; chart titles are always centered.
• Notes pinned to a uniform bottom position (margin-top:auto), same on every page.
• Nothing overlaps the footer line. .pad reserves bottom clearance.
• Fill by composition, not by stretching — <30% whitespace. Sparse prose spreads (space-evenly); sparse tables keep their uniform row height (header 34px / rows 52px, deck-wide) and take an extra= stat strip or more researched rows. Never balloon an element to hide thin content.
• No mid-word truncation. Use the engine's clipw() (word boundary + "…"), never raw [:n].
• Images: clean cutout on pure white, no shadow, mix-blend-mode:multiply to blend; every image gets an italic caption. Cover/divider images are the only ones on navy.
• Flat only: no drop shadows, no bevels (navy cover/divider excepted).
• Self-verify by rendering the PDF, not by trusting the browser tab (an already-open tab won't reload — tell the user to hard-refresh with Cmd+Shift+R).

Quality bar

Match AND exceed a reference McKinsey report on both content density and visual polish. Prose pages use "bold theme + flowing paragraph", not terse bullets. Be honest about assumptions — flag the load-bearing judgment calls rather than hiding them; that honesty is what reads as senior.

Two final gates before delivery:

• Elevator test — the §0 governing thought + thesis page must sell the recommendation in 30 seconds, standalone. If it needs the rest of the deck to make sense, it isn't the answer yet.
• Horizontal logic — re-run the headline read-through on the finished deck (QC §C); page edits during build often break the essay.