tekkiiiii/the-agency/deck-narrative

Claude Code and agent skill by tekkiiiii/the-agency

Trust tier 2
  • community
LLM review pending
Source check pending

Checking that the SKILL.md file is still in the repository.

Use to review or fix a deck's argument, not its look: \"does this deck make sense\", \"is this an action title\", \"Minto pyramid\", \"one message per slide\", \"decision slide\", \"check deck fonts\". Rules as pass/fail checks; ships scripts/audit_deck.py for .pptx. Co-load with a visual deck skill.

Install

npx vskill@latest install tekkiiiii/the-agency/deck-narrative

SKILL.md

The same SKILL.md was published earlier as tekkiiiii/the-agency/deck-narrative.

deck-narrative

Deck-building skills in this system (e.g. strategic-deck, marp, or whichever visual/deck skill your setup uses) are all visual systems — CSS classes, palettes, fonts, slide-type names, build mechanics. None of them say anything about whether the content is any good: whether the title makes a claim, whether the deck argues top-down, whether a table should have been a chart. This skill is the missing layer underneath all of them. It is a reasoning discipline, not a template.

Read this first: architecture vs visual system

LayerDecidesOwned by
Architecture (this skill)What goes on the slide, in what order, and whether it makes an argumentdeck-narrative
Visual systemWhat it looks like — color, type, layout, motionstrategic-deck / marp / whichever visual skill your setup uses

Co-load, don't choose. When building or reviewing a deck, load deck-narrative alongside whichever visual skill applies (your project's routing convention decides which visual skill to co-load). This skill never touches CSS, pptxgenjs calls, or slide-type names — it decides the titles, the sequence, and whether a slide should be a table or a chart, then hands the visual skill the content to lay out. Running a visual deck skill without this one is how decks get built by instinct: correct fonts, no argument.

The one narrow exception (Rule 9): which brand font to use, its weight, its pairing — those stay the visual/brand skill's job (whichever brand skill is co-loaded). But whether the shipped .pptx FILE actually contains that font, versus an Office default that survived because a rendering pipeline substituted a lookalike glyph, is not an aesthetic judgment — it's a file-integrity check a visual review structurally cannot make (see Rule 9 below). That single check lives here because it needs to read the .pptx XML, not because deck-narrative has an opinion on typography.

Provenance and confidence

Source: NotebookLM deep research, 61 web sources — McKinsey/BCG/Bain deck conventions, Barbara Minto's pyramid principle, Michael Alley's assertion-evidence structure, Edward Tufte's data-ink principle. Full source: internal research archive (not included in this public skill mirror).

  • Structural rules (action titles, one-message-per-slide, top-down

sequencing, horizontal/vertical logic) — HIGH confidence. Multiple independent sources converge.

  • Exact numeric thresholds (word counts, pt sizes, slide counts, minutes

per slide) — MEDIUM confidence. These are conventions that vary by source, not laws. Present them as defaults with stated reasoning. Never treat a threshold miss as a hard gate that blocks shipping a deck — flag it, explain why the convention exists, let the human/agent decide.


Rule 1 — Action titles vs topic titles

The rule: a slide headline must make a complete, active claim, not name a category.

Checks:

  • Contains a verb — not just a noun phrase.
  • ≤15 words.
  • ≤2 lines.
  • "So what" test: the title states the strategic implication of the

slide's content, not just a label for it. (Judgment — see below.)

  • Layout: left-justified, 30-32pt. Under 28pt is a text-stuffing red flag

(MEDIUM-confidence convention — a slide with a 24pt title isn't broken, but it's a signal the title got squeezed to fit too much).

Before/after:

  • Before (topic label): "Launch Readiness"
  • After (action title): *"Launch Is 6 Weeks Behind Schedule — Three Decisions

Required by Friday to Recover the Timeline"*

Worked example (anonymized, from a real audit) — a defence-sector trade-show deck audit (see scripts/audit_deck.py output below): slide 8's title "Từ hợp tác toàn cầu đến làm chủ công nghệ" ("From global cooperation to technology mastery") is a topic label — a "from X to Y" bridge phrase with no independent verb clause and no claim. It names the section, it doesn't argue anything. Compare slide 14: "Kênh duy nhất tăng tiếp cận với chi phí media bằng 0" ("The only channel that grows reach at zero media cost") — a verb ("tăng"/"grows"), a claim, a number backing it up.

Mechanical vs judgment: the verb test and word/line counts are mechanical — audit_deck.py checks them. The "so what" test (does the claim carry implication, not just description) is judgment — flag candidates, don't auto-fail them.


Rule 2 — One message per slide

The rule: each slide conveys exactly one distinct insight.

Checks:

  • The "so what" element test: point at any chart, shape, or text block on

the slide. If it doesn't directly prove the action title, delete it.

  • The 5-second test: a viewer grasps the core point within 5 seconds of

first seeing the slide.

  • The 10-second test: a reader can explain the slide's purpose and proof

back to you in 10 seconds.

  • The 90% whitespace rule: if the canvas is ≥90% full, it's overloaded —

split it into two slides.

Mechanical vs judgment: none of this is mechanically checkable from a .pptx file alone — it requires reading the slide the way a viewer would. This is human/agent judgment, always. audit_deck.py cannot run the 5-second test; a critic agent with the rendered slide can.


Rule 3 — Minto pyramid / top-down sequencing

The rule: present arguments top-down — conclusion first, supporting pillars second, granular evidence third. Never save the recommendation for a "surprise ending."

Checks:

  • Conclusion first: the recommendation appears on slide 2, not buried at

the end.

  • Title slide (slide 1): exactly three elements allowed — presentation

title (<10 words), client/company logo, date. Nothing else.

  • The five-minute window (slides 1-3): these three slides completely

outline the entire recommendation. A reader who only sees slides 1-3 should know the whole argument.

  • The bold-bullet summary (slide 2): use bold-bullet format; the bolded

phrases alone must tell the entire story when skimmed.

Worked example (same anonymized deck): puts the single committed KPI ("Reach is the only committed KPI") on slide 2, immediately after the cover — not a section-by-section build toward a reveal. That's the pattern to replicate.

Mechanical vs judgment: "is slide 2 actually the conclusion" and "do slides 1-3 outline the whole recommendation" require reading content and comparing against the deck's actual thesis — judgment, not something audit_deck.py can verify from shape geometry.


Rule 4 — Pacing and text density

The rule: align slide count and formatting to the meeting length, to prevent live cognitive overload.

Checks (MEDIUM-confidence conventions, present as defaults):

  • 20-minute slot → target ~10 slides.
  • 30-minute slot → 10-15 slides (max 15-25).
  • Board meeting → 10-15 slides.
  • Reserve the last ~10 minutes of a 30-minute slot entirely slide-free, for

Q&A/discussion.

  • Pacing: 1-2 minutes per slide.
  • On-slide limits: max 5-7 bullets, max 6-8 words/line, max 2-3 lines of body

text. Apply the 5×5 rule (5 words/line, 5 lines/slide) or 7×7×7 as house convention allows.

Mechanical: slide count and word-per-title checks are mechanical (audit_deck.py reports both). Bullets-per-slide and words-per-line on body content would require parsing every text box on every slide — not currently in the script; extend it if a deck's body density becomes a recurring problem.


Rule 5 — Horizontal and vertical logic

The rule: verify narrative continuity across the deck (horizontal) and empirical support within each slide (vertical).

Checks:

  • The read-through test (horizontal): strip away everything but the

titles, read them in sequence. If they don't read as a logically flowing, persuasive essay — the deck's "narrative spine" — it fails.

  • The dual-coding test (vertical): strict 1-to-1 correlation — the visual

in the slide body must immediately prove the claim in the title.

Mechanical vs judgment: audit_deck.py prints the full title sequence in order specifically so a human or agent can run the read-through test — the script cannot judge whether titles "read as a persuasive essay," but it removes the friction of extracting them by hand. The dual-coding test (does the body prove the title) needs the slide content, not just its shape — judgment.


Rule 6 — Tables vs charts

The rule: choose the data format for the audience's cognitive task, not visual variety.

Checks:

  • Use a TABLE when: presenting precise values to look up/compare, sparse

data, ≤20 numbers, or when a zero-baseline would compress/distort variation.

  • Use a CHART when: showing trends, variations, or relationships. Match

chart type to comparison goal:

  • Trend over time → line chart
  • Categorical ranking → bar/column chart (pie capped at 6 categories)
  • Part-to-whole → stacked bar or Mekko (cap at 5 segments)
  • Cumulative walk → waterfall chart

Mechanical: audit_deck.py counts tables vs charts per slide and deck-wide (via shape.has_table / shape.has_chart). It does NOT judge whether a given table should have been a chart — that requires knowing the underlying data shape and the audience's task, which is judgment. A deck-wide total_charts: 0 with several tables holding trend data (numbers changing across time periods) is a real signal worth a human/agent look — the script surfaces the count, not the verdict.

Hand-off: when a table should become a chart, or a new chart is needed, use whichever chart-rendering skill or tool your setup provides to produce a native, editable chart (e.g. a pptxgenjs addChart() call, or an SVG-as-PNG embed for chart types with no native PPTX equivalent: waterfall, heatmap, bullet). If the right chart type isn't obvious, pick it by data relationship + category count + audience task (see Rule 6's checks above) before rendering. deck-narrative decides that a chart is needed and what kind in principle (Rule 6 above); your chart-rendering tool renders it.


Rule 7 — Decision / approval-request slides

The rule: action slides must clearly define options, stakes, and next steps to force an outcome.

Checks:

  • Visual placement: the requested decision sits in a distinct callout box

at the bottom of the slide.

  • Grammar standard: bullets always begin with an active verb (Grow,

Reduce, Improve, Increase, Target).

  • Approved structural frameworks — audit against one of these three:
  1. Challenge > Options > Decision — problem, paths, recommendation.
  2. Goal > Obstacle > Strategy > Action — objective, barriers, approach,

direct steps.

  1. Insight > Action > Outcome — finding, proposal, expected return.
  • Credibility: every data point cites a source line at 8-9pt.
  • Pre-wire verification: confirm key recommendation slides were shared

individually with stakeholders before the meeting — eliminate surprises.

Worked example (same anonymized deck): slide 22, "Đề nghị ban lãnh đạo / người phê duyệt phê duyệt các nội dung sau" ("Requesting the board / the approver approve the following") — an active-verb approval ask. Slide 20's title ("Người trình bày → Người phê duyệt → CEO + Chủ tịch" / "Presenter → Approver → CEO + Chairman") is the governance-chain slide feeding into it — worth checking it isn't standing in for the actual decision framework above.

Mechanical vs judgment: active-verb-bullet detection and callout placement could be scripted (not yet in audit_deck.py — extend if this becomes a recurring gap). Which of the three frameworks a given slide fits, and whether pre-wiring actually happened, are judgment calls outside what any script can see in a .pptx file.


Rule 8 — Reader-only content (no process meta in the .pptx, and no

spoken material on the slide)

This is the start of the file. Read the full SKILL.md on GitHub.

Security

  • Security scanPassed, 100 of 100Pattern checks on the skill's files.
  • Model reviewNot runA model review is scheduled for the next rescan.

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
3
7-day trend score
7

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/tekkiiiii/the-agency/deck-narrative/badge)](https://verified-skill.com/skills/tekkiiiii/the-agency/deck-narrative)

Something wrong with this skill or its page? Tell us.