agentleFS
mikel-codemikel-code/claude-skills/mckinsey-deck/SKILL.md

mckinsey-deck

Skill0 starsChanged 3 months agoMITView on GitHub

Create or iterate on slide-deck content in Markdown applying McKinsey criteria (Pyramid Principle, action titles, MECE, so-what, quantification with sources). User provides context files and a deck goal; skill proposes an outline, then produces one slide at a time with feedback loops. Output is a deck.md file — content only, no visual references.

The file

Show its rules and text

We split this skill into the rules it gives an agent

17 rules. About 0 are common, 12 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈12

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every deck is a single Markdown file (deck.md in the working directory, unless the user specifies otherwise).

    line 15
  2. Very specific:

    Structural rule: Every slide is bounded by ---.

    line 17
  3. Distinctive:

    Update the manifest comment each time.

    line 62
  4. Distinctive:

    Before proposing the outline, before writing each slide, and before the final pass — check every item:

    line 73
  5. Distinctive:

    Read every context file and source the user provides.

    line 96
  6. Distinctive:

    Do not produce output yet.

    line 103
  7. Distinctive:

    After the outline, write a short paragraph (≤5 sentences) that:

    line 119
  8. Distinctive:

    Stop. Wait for the user's feedback: Do not write any slide body.

    line 125
  9. Very specific:

    Create deck.md (or the user-specified path) with the manifest comment.

    line 134
  10. Very specific:

    Write the # cover slide as the deck's first --- block: deck title, optional subtitle, optional audience/attribution, and date.

    line 135
  11. Very specific:

    After the slide markdown, add a > Lens review: block listing how each applicable lens is satisfied, and flagging anything marginal.

    line 138
  12. Distinctive:

    After the last slide is approved:

    line 147
  13. Distinctive:

    Verify traceability: every figure, code, and name is consistent throughout.

    line 150
  14. Distinctive:

    Verify MECE across sections.

    line 151
  15. Distinctive:

    Tier 3 — Unknown / unverifiable: If a claim is covered by no source and is not first-party knowledge either, flag it explicitly, inline:

    line 163
  16. Distinctive:

    Never pass off an untraceable third-party claim as fact: If a number purports to be public but you cannot trace it, flag it (Tier 3) — do …

    line 165
  17. Distinctive:

    All McKinsey criteria, principles, examples, and checklists encoded in this skill are consolidated in:

    line 173

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

McKinsey Deck Builder

Produce slide-deck content in Markdown following McKinsey's communication discipline. Scope is content only: action titles, body copy, footnotes, and sources. Never write chart types, colors, icons, layout instructions, or any visual reference.


Output format rules

Every deck is a single Markdown file (deck.md in the working directory, unless the user specifies otherwise). Slides are separated by ---.

Structural rule: Every slide is bounded by ---. Each --- block contains exactly one heading — a single # (cover slide, deck's first block only), ## (section-intro slide, no body), or ### (content slide, body follows). Never more than one heading in the same block.

Heading levelSlide typeRules
#Cover slideAlways the deck's first --- block, and the only place a # may appear. Format: # <Deck title> followed by an optional subtitle line, an optional audience/attribution line, and a date. No action title, no body bullets, no Sources: block. Exactly one per deck.
##Section-intro slideFormat: ## NN — Section title where NN is the section index (01, 02 …), not the slide index. No body. No ### in the same block.
###Content slideFormat: ### Action title that states the conclusion. Body content follows. Must be in its own --- block, separate from any ##.

Canonical layout opening with the cover, then one section with two content slides:

# <Deck title>

<Subtitle — what the deck delivers>

<Audience / attribution> · <Date>

---

## 01 — Situation

---

### Action title for the first content slide

<!-- comparison: time-series -->

Body…

---

### Action title for the second content slide

<!-- comparison: none -->

Body…

---

## 02 — Complication

Deck file management:

  • First slide: create deck.md with a manifest comment at the top (<!-- deck: <goal summary> | slides written: N of M -->), then write the cover slide and slide 1. N of M counts every --- block including the # cover and ## section-intro slides.
  • Subsequent slides: Edit the file to append the new slide (do not rewrite previously approved slides).
  • Update the manifest comment each time.

Comparison tag (content slides only):

  • Immediately after the ### action title, emit a hidden, non-rendering tag on its own line: <!-- comparison: <type> -->.
  • <type> is the Zelazny comparison the message makes: component (share of a whole), item (ranking), time-series (change over time), distribution (spread across ranges), correlation (relationship between two variables), or none (qualitative/conceptual — not chartable).
  • This is a content-level signal for the downstream design step — never a visual instruction, and it never appears on the rendered slide. Cover (#) and section-intro (##) slides take no tag.

McKinsey lenses (apply at every step)

Before proposing the outline, before writing each slide, and before the final pass — check every item:

  1. Pyramid Principle (Minto) — Answer-first at deck and slide level. The conclusion is the first thing stated, not the last.
  2. Day-one answer — The deck opens with the answer and its 3-4 supporting pillars before any supporting detail.
  3. Action titles — Every ### heading is a full declarative sentence stating the conclusion (not a topic label). Reading only the ### titles in order must tell the complete story: the horizontal flow test.
  4. MECE — The sections and slides cover the whole problem space with no overlap. Run the MECE test on every grouping.
  5. Quantification with source per figure — Every number or statistic carries a specific value, not a range or vague qualifier. Public/third-party figures also carry an inline reference marker [N] immediately after the figure, with the full Source: <name>, <date> listed in the slide's end-of-slide Sources: block (reuse the same [N] when one source backs multiple claims). Figures that originate from the user's own internal/first-party material carry no citation — state them plainly with no [N] and no Sources: entry (see Reference validation).
  6. One idea per slide — No two independent arguments on the same content slide. If two ideas are needed, create two slides.
  7. Synthesis, not summary (the "So what?" test) — Every content slide must carry the client's implication: what does the evidence mean for them, and what decision does it call for? Embed the implication in the action title and/or a closing bullet written as a natural conclusion. Do not append a labeled **So what:** footer to every slide — that is a crutch that signals the implication was not properly embedded in the body. Use an explicit implication call-out only when the slide demands a specific decision from the audience.
  8. Low text density — ≤40 words per body block. Use bullet points, short clauses, or markdown tables. No paragraphs.
  9. Traceability and consistency — Figures, codes, names, and acronyms are identical across every slide in the deck. Verify before finalizing each slide.
  10. SCQA opening — The first slides of the deck follow Situation → Complication → Question → Answer.
  11. Hypothesis-driven — Each section (##) corresponds to a key hypothesis or issue-tree branch, not a generic topic.
  12. Recommendation separate from risks — The recommendation slide states the answer with conviction, using imperatives. Caveats, constraints, and risks appear on a dedicated slide.
  13. No fluff — No warm-up bios, no suspense arcs, no hedging language ("it could be argued that…"), no decorative phrases.
  14. Chart-readiness and context — Decide whether the slide's message is a quantitative comparison or qualitative, and record it in the <!-- comparison: <type> --> tag. A chart is warranted only when the message turns on a comparison; a lone figure is a number, not a chart. When the message is a comparison, it must carry its comparator — target, prior periods, peer, or threshold (lens applies principle #14, Benchmarking and Context, to exhibits). If that comparator is absent from the sources: ask the user for it when the comparison is the slide's core, load-bearing message; otherwise drop a ⚠ Missing context — <what is needed> flag. Never present a single-value "chart" or a comparison stripped of its benchmark.

Workflow (mandatory; follow exactly)

Step 1 — Ingest

Read every context file and source the user provides. Read the deck goal carefully. Think hard before responding:

  • What is the central question the deck must answer?
  • What is the day-one answer hypothesis?
  • What is the SCQA frame for the opening?
  • What are the 3-5 key hypotheses or drivers the deck must prove?
  • What upstream sources (regulations, statutes, URLs, datasets, named studies, agency authorizations) do the context files themselves cite? Note these now — they are the Tier 1 citations for slides, not the filenames.

Do not produce output yet. Proceed to Step 2.

Step 2 — Propose outline

Output only the outline — no slide bodies. Format:

## 01 — <Section title>
### <Action title (full sentence)>
### <Action title (full sentence)>

## 02 — <Section title>
### <Action title (full sentence)>
...

After the outline, write a short paragraph (≤5 sentences) that:

  1. States the day-one answer hypothesis.
  2. Explains the SCQA frame.
  3. Confirms the horizontal flow test passes (quotes the 3-5 most critical ### titles that carry the story).
  4. Confirms the structure is MECE.

Stop. Wait for the user's feedback. Do not write any slide body.

Step 3 — Iterate on outline

Apply the user's changes. Repeat Step 2 output format + the short paragraph. Continue until the user explicitly approves the outline.

Step 4 — Write slide 1

After outline approval:

  1. Create deck.md (or the user-specified path) with the manifest comment.
  2. Write the # cover slide as the deck's first --- block: deck title, optional subtitle, optional audience/attribution, and date. Derive the title from the deck goal; keep it factual, no action-title phrasing. Then write slide 1, applying all 13 McKinsey lenses above.
  3. Validate every reference in the slide (see Reference validation below).
  4. Format check — confirm the cover is the first block and the only #, and each slide's --- block contains exactly one heading at the correct level (# cover, ## section-intro, ### content). If a block has more than one heading, fix before proceeding.
  5. After the slide markdown, add a > **Lens review:** block listing how each applicable lens is satisfied, and flagging anything marginal. Include these three lines: Format: one heading, correct level, own --- block — ✓; Sources: public upstream cited; internal/first-party claims left uncited; unverifiable claims flagged — ✓; and Comparison/context: <type> tagged; comparator present, or asked/flagged when missing — ✓.
  6. Stop. Wait for the user's feedback.

Step 5 — Iterate and advance

Apply the user's feedback to the current slide (use Edit tool to modify only that slide). Then write the next slide. Repeat Step 4 procedure (write + lens review + stop) until all slides in the approved outline are complete.

Step 6 — Final pass

After the last slide is approved:

  1. Re-read deck.md in full.
  2. Confirm the deck opens with exactly one # cover slide and that the executive-summary slide immediately follows it. Run the horizontal flow test on all ### titles.
  3. Verify traceability: every figure, code, and name is consistent throughout.
  4. Verify MECE across sections.
  5. Comparison-tag audit — for every content slide, confirm a <!-- comparison: <type> --> tag is present; for each non-none tag, confirm the slide body holds its comparator (second series, prior periods, peer, target, or threshold). Where a comparator is missing, ask the user to supply the benchmark; if they cannot, downgrade the tag to none (only with their confirmation) and ensure the slide carries a ⚠ Missing context note. This makes deck.md authoritative before the design step, so deck-design-prompt can trust the tags.
  6. Report any remaining issues or flagged references.

Reference validation (mandatory for every slide)

For every figure, statistic, quote, name, attribution, date, or external claim included in a slide apply this hierarchy:

  • Tier 1 — Public upstream source. If the claim traces to a public, third-party-verifiable source (regulation, statute, URL, dataset, study, agency authorization, named third party), cite that upstream source — not the filename of the internal file it came through. Example: a context file quoting "per Ley 10/2010 Art. 17, blocking is required" → Source: Ley 10/2010, Art. 17, not Source: Informe AML.md. Never name an internal context file as a via/intermediate source.
  • Tier 2 — Internal / first-party origin. If the claim originates from the user's own non-public material — an internal file, proprietary observation, in-house analysis, the user's own thinking, or first-party company data — include it with no citation at all: no [N] marker and no Sources: entry. It is presented as the author's own assertion; a private filename adds no credibility to the audience and is routinely stripped from the deck. Do not invent a public source to dress it up.
  • Tier 3 — Unknown / unverifiable. If a claim is covered by no source and is not first-party knowledge either, flag it explicitly, inline:

    ⚠ Unverified — please confirm: "<exact claim>" (suggested source if any)

  • Never pass off an untraceable third-party claim as fact. If a number purports to be public but you cannot trace it, flag it (Tier 3) — do not cite the internal file in its place.

Reference format: place an inline marker [N] immediately after each Tier 1 figure or claim. Collect those references in a Sources: block at the foot of the slide, one per line as [N] Source: <name>, <date>; reuse the same number when one source backs multiple claims. Omit the Sources: block entirely when a slide has no Tier 1 citations — an all-internal slide carries none.


Reference file

All McKinsey criteria, principles, examples, and checklists encoded in this skill are consolidated in:

<skill-dir>/references/mckinsey-principles.md

Re-read that file at invocation time whenever a deeper reference on any criterion, example, or checklist is needed.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.