Skip to content

JPeetz/agent-skills / skills/documentation-content/presentation-prompt-pipeline

presentation-prompt-pipeline

Verified

Use this skill when building or repairing a presentation with an AI model — the six staged prompts (blueprint → structure → story frame → visual direction → content → build) that turn an idea into a deck with one message, each stage producing an artifact and a checkable gate, with prompt templates, eight named failure modes, and a deck_lint.py script that mechanically enforces short titles, tight bullets, speaker notes, a two-font and three-colour ceiling.

Presentation Prompt Pipeline

Six staged prompts that take a deck from idea to presentable, in the order the work actually has to happen: message → outline → story → visual direction → copy → file. Each stage produces an artifact the next stage consumes, and each has a gate you can check before spending the next prompt. This skill is the planning and review layer; the mechanical creation of the .pptx is plain python-pptx.

When to Use

  • Building a deck with a model from a brief, a document, or a pile of notes
  • A deck exists but reads as a wall of bullets, or says nothing memorable
  • Pitching, teaching, or reporting where the ask on the last slide matters
  • Reviewing someone else's deck against a repeatable standard
  • Don't use for: decks whose content is fixed by a template or mandate (drive python-pptx directly), format conversion, or as a substitute for rendering the slides and looking at them

The Six Stages

#StageProducesGate before moving on
1BlueprintOne message, audience, 3 emotional angles, slide flow, target slide countThe deck's single message fits one falsifiable sentence, and you can name who it changes
2StructureSlide-by-slide outline: one-line purpose and intended feeling per slideEvery slide has a purpose line; no slide exists "for context"
3Story frameHook → Problem → Insight → Solution → Proof → Takeaway beat sheet mapped onto slidesAll six beats present; the last slide carries exactly one action
4Visual directionEmotional visual intent, chart-for-data choice, icon/grid for lists, 3-colour palette, font pairingPalette and fonts are named with hex/font values and cover every slide
5ContentTitles ≤6 words, 3–5 bullets ≤10 words, one speaker note per slide, audience-matched tonedeck_lint.py passes on the draft text
6BuildThe .pptx: bold titles, tight bullets, notes on every slide, clean flowdeck_lint.py passes on the real file; rendered slides reviewed by eye

Prompt Templates

Each prompt names its input and the shape of its output — that is what makes the output checkable rather than merely plausible.

1 — Blueprint

You are a presentation strategist. From the material below, give me: (a) the ONE message in a single sentence, phrased as a claim, not a topic; (b) the audience and the three things they care about most; (c) three emotional angles I could take and the risk of each; (d) a slide-by-slide flow in ≤12 steps; (e) the ideal slide count and why that number. Be specific — no generic advice. Material: {brief}

2 — Structure

Using this blueprint, write a slide-by-slide outline. For each slide: a one-line purpose ("this slide makes the audience ___"), what they should feel, and the single idea it carries. Flag any slide that repeats another's job. {blueprint}

3 — Story frame

Rewrite the outline into this six-beat story frame: Hook (grab attention), Problem (the pain), Insight (the surprise), Solution (what to do), Proof (one statistic or example), Takeaway (one action). Map each beat onto slide numbers, and say in one line what the audience should feel at each beat. {outline}

4 — Visual direction

For each slide, specify: the emotion the visual must carry, whether the data is better shown as a bar, line, pie, or a single big number and why, whether lists become icons or a grid, one 3-colour palette with hex codes (background, ink, accent), and one heading/body font pairing available in standard Office or Google fonts. No stock-photo descriptions; describe the composition.

5 — Content

Write the final copy per slide: a title of ≤6 words, 3–5 bullets of ≤10 words each (no sentences), and one speaker note of 2–3 sentences that says what to say out loud — not a restatement of the bullets. Match this tone: {tone}. Audience: {audience}. Keep the Takeaway slide to a single action.

6 — Build

Produce a python-pptx JSON spec for this content: 16:9, title+content layouts, the palette and fonts above, speaker notes included per slide. Then the deck is built and linted — not hand-assembled.

How to Run

Stages 1–5 are prompts: run each in a fresh context with the previous stage's artifact pasted in, not as one mega-prompt. Stage 6 builds the file and then lints it:

uv run --with python-pptx scripts/deck_lint.py deck.pptx

The linter is the only script here. It reads a .pptx, prints JSON or a human line per finding, and exits 0 clean, 1 with findings, 2 unreadable file. Tune the ceilings per deck:

uv run --with python-pptx scripts/deck_lint.py deck.pptx \
  --max-title-words 8 --max-bullets 6 --max-bullet-words 12

Rules it enforces by default: title ≤6 words, ≤5 bullets per slide, ≤10 words per bullet, speaker notes on every slide, ≤2 font families, ≤3 explicit text colours.

Quick Reference

SymptomWhere the process broke
Message is fuzzyStage 1 accepted a topic ("Q3 review") instead of a claim
30 slides of bulletsStage 2 outline has no purpose lines, so nothing could be cut
Every slide feels the sameStage 4's emotional intent was never specified
Titles are sentencesStage 5's ≤6-word rule, enforced by the linter
Data slide is unreadableStage 4: one chart type per idea, one number per chart
Deck "done" but not deliveredNotes missing, or the last slide has no single ask

Pitfalls

  1. One mega-prompt. Asking for message + outline + copy + design at once produces a generic deck, because the model optimises every constraint at once. The stages exist so each constraint gets attention.
  2. A blueprint that names a topic, not a claim. "Q3 review" is not a message; "Churn doubled and ownership fixes it" is. Stage 1's output must be falsifiable.
  3. Bullet walls dressed as structure. 5 bullets × 12 words reads as prose. The ≤10-word rule forces one idea per bullet.
  4. Decoration before storyline. Picking a theme or template first locks you into layouts that fight the story; stage 4 comes after stage 3 for that reason.
  5. Speaker notes that repeat the bullets. Notes are the spoken track; if they duplicate the slide, the deck is a document, not a presentation.
  6. Palette creep. Every extra accent colour dilutes the one that matters; the linter caps explicit text colours at three by default.
  7. Trusting the outline instead of the rendered slide. Overflow, overlap and contrast problems exist only in the render — always look at the slides.
  8. A Takeaway slide with three asks. One action. Three asks means the deck has three messages and none of them lands.

Verification

  • deck_lint.py exits 0 on the final file
  • Slides rendered to images and reviewed — no overflow, no contrast failure, no orphaned bullet
  • Stage 1's one-sentence message appears near the front, and the Takeaway slide repeats the action rather than the message
  • Every slide traces to a stage-2 purpose line; none is "context"

Reference

  • references/six-prompts-cheatsheet.md — transcription of the source infographic (Cyberman AI, "How to build a presentation using AI: 6 powerful prompts"), including each stage's checklist items.
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall