TL;DR
Build the outline in a chat model (Claude, ChatGPT, or Gemini) first, then hand the approved outline to a slide tool. A model writes the argument; it does not know which point will offend your VP. Feed it audience, one-sentence message, three points + evidence, and time slot. Demand slide titles that read as a coherent argument when scanned, plus a CTA that names who / what / by when. Then generate slides in Gamma, Claude Design, or Gemini Canvas and replace every AI-invented number with your real one.
The task
You have 20 minutes to present a project, an idea, or a recommendation to a mixed audience: your manager, a couple of execs, peers who already know the work, and a stakeholder who knows nothing. You have plenty of material and zero structure. The goal is a 10-slide outline with one message, three supporting points anchored in evidence, slide titles that read coherently if a stranger scans them, and a final slide with a concrete call to action.
The trap most people fall into is generating the slides first. A pretty deck built on a muddy argument is still a muddy argument. Lock the outline, then make it look good.
When AI helps — and when it does not
AI is genuinely good at three things here: distilling messy notes into a slide sequence, suggesting visuals, and drafting speaker notes. It is poor at organizational politics. If your recommendation contradicts a sponsor’s prior public position, the model will not flag it — it has no map of who said what in last quarter’s review. Read the deck against your stakeholder map before sharing, and pre-brief anyone who will be surprised.
It also will not know which of your numbers is real. Every model can and will fabricate a plausible statistic to fill a chart. Treat each figure it produces as a placeholder until you replace it with your own source.
What to feed the AI
- Audience + their primary goal (decide, align, learn, fund)
- One-sentence message you want them to leave with
- Three supporting points with the evidence behind each
- Time slot (20 / 30 / 45 min)
- Format constraints (in-person, hybrid, slides only, async Loom)
- Anything explicitly out of scope or pre-decided
Thin input produces generic output. The single biggest quality lever is the one-sentence message — if you cannot write it in one line, the model cannot structure around it.
Copy-ready prompt
Outline a 10-slide work presentation.
Audience: [list of roles]
Audience primary goal: [decide / align / learn / fund]
One-sentence message: [line]
Time slot: [minutes]
Format: [in-person / hybrid / Loom]
Out of scope or pre-decided: [list]
Supporting point 1: [line] — evidence: [data / quote]
Supporting point 2: [line] — evidence: [data / quote]
Supporting point 3: [line] — evidence: [data / quote]
For each of 10 slides, return:
- Title (under 8 words, a full sentence not a topic label)
- One supporting bullet
- A visual idea (chart / image / table)
- 30-word speaker note
- A "cut if short on time" flag for slides that can be sacrificed
Structure:
- Slide 1: hook tied to audience pain
- Slides 2-4: point 1 (build -> evidence -> implication)
- Slides 5-7: point 2 (same structure)
- Slide 8: point 3 (single slide if shorter)
- Slide 9: risks / open questions
- Slide 10: concrete call to action with owner and timeline
Concrete CTA only. No "let's keep the conversation going."
Do not invent statistics. Where a number is needed, write [INSERT REAL FIGURE].For decision-seeking meetings, add: “Insert a recommendation slide before the CTA, listing the 2 alternatives I considered and rejected and one line on why.”
That last instruction — [INSERT REAL FIGURE] — is the difference between a deck you can ship and one that quietly lies to your execs. The 2026 chat models (GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro) will all happily generate “37% improvement” if you let them.
Which tool: outline first, then slides
Two separate jobs. The chat model writes and refines the argument. The slide tool turns the approved outline into a deck. Pricing and capabilities below are current as of June 2026.
| Tool | Best at | Free tier | Paid entry | Outline-first? |
|---|---|---|---|---|
| Claude (Opus 4.7) | Tight long-form structure; HTML/PPTX decks via Claude Design | Limited Sonnet 4.6 | Pro $20/mo | Yes — argue, then “Export as PPTX” |
| ChatGPT (GPT-5.5) | All-round drafting; Canvas editing | $0 (tight limits, US Free has ads) | Plus $20/mo | Yes, in Canvas |
| Gemini (3.1 Pro) | Pulls from NotebookLM; one-click to Google Slides | $0 | Google AI Pro $19.99/mo | Yes, in Gemini Canvas |
| Gamma | Fastest finished-looking deck from a topic | 400 one-time credits, watermark | Plus $12/mo | Yes — edit outline before generate |
Notes from June 2026:
- Claude Design (launched April 17, 2026) pairs Opus 4.7 with an HTML canvas and exports six ways, including PPTX, standalone HTML, and Send to Canva. Best when you want a real PowerPoint file out the door.
- Gemini Canvas generates a full deck and exports to Google Slides in one click (rolled out broadly April 11, 2026). Layout control at generation is limited — it picks layouts by content type — and image quality is inconsistent, so plan on swapping in brand assets.
- Gamma is the fastest from a cold start. Its free plan is 400 one-time credits (generating a deck costs ~40, each added card ~5) and leaves a watermark; Plus is $12/mo with 1,000 refreshing monthly credits, Pro $25/mo adds analytics and custom domains.
Whichever you pick: budget 30–50% of the time you “saved” for post-generation editing. Every 2026 AI slide generator gives you a first draft, not a final deck.
How to check the output is usable
- A stranger reading only the titles in order gets the whole argument
- Each point is built on evidence, not opinion
- The CTA names who, what, and by when
- Risks / open questions appear before the CTA, not after
- The cut-if-short slides are genuinely cuttable
- Every chart carries your number, not a model-invented one
Common mistakes
- One-message-per-slide forgotten — slides with three points blur into one
- No call to action — meetings without a CTA end and nothing happens
- Background eating slides 1–4 — exec audiences walk away by slide 3
- Visual ideas described in words only — “a chart here” never becomes a chart at the meeting
- Letting AI invent data — every figure needs your source behind it
- Politically risky points with no sponsor pre-brief — a bad surprise in the room
FAQ
Should I build the outline and the slides in the same tool? You can, but separating them produces a better argument. Write and pressure-test the outline in a chat model, get it approved by a peer, then generate slides. Generating slides first locks you into a structure you have not actually validated.
20 vs 45 minutes — how many slides? 20 minutes is 5–7 slides; 45 is 10–12. Budget Q&A separately and keep one or two slides flagged “cut if short.”
Which tool exports a real .pptx file? As of June 2026, Claude Design exports PPTX directly, and Gamma and Gemini export to their own formats / Google Slides. ChatGPT can produce a .pptx via its agent mode but the output is aesthetically plain.
Should I rehearse? Yes — once timed, once for energy. Twice is enough for a 20-minute talk.
What if the audience hijacks the meeting? Use the cut-if-short flags to drop slides live and still land the CTA. That is the whole reason you flagged them.
Is the AI’s data trustworthy? No. Treat every number it produces as a placeholder. Models fabricate plausible statistics to fill charts; replace each one with a figure you can cite.
Related
- PPT outline AI — broader presentation outline workflow
- PPT outline prompts — additional outline prompt variants
- Presentation speaker notes prompts — speaker notes after the outline
- AI generate PPT tutorial — outline → finished slides workflow
- Self introduction — your opening 60 seconds
- Storytelling outline — story-shaped presentations
- Manager update email AI — the written companion to the talk
Tags: #AI writing #Job search #Workflow