TL;DR
Feed the model one or two bullets per slide (not the raw deck), name the three questions your CEO will ask, and cap the output at 4 ranked bullets, 100-120 words, each ending in an action verb. The copy-ready prompt below does exactly that. As of June 2026, any of GPT-5.5, Claude Sonnet 4.6, or Gemini 3.1 Pro handles this in one shot; the only reason to upload the whole deck is if it runs longer than your plan’s context window — see the model table.
The task
You have a 12-slide management report — financials, KPIs, market notes, risk register. You need a one-screen exec summary the CEO can read while walking into the meeting. This is a compression job, not an analysis job, and that distinction decides how you prompt.
When this is the right job for AI
- The full report is already finished and accurate.
- You can extract the page-by-page bullets (or paste the slide notes) without copying sensitive raw data.
- You want compression, not new analysis. The model is excellent at ranking 12 pages of bullets and dropping the bottom 8; it is bad at inventing a forward view your deck never stated.
What to feed the model
- 1-2 bullet points per slide (do not paste full slides).
- The 3 questions the CEO is most likely to ask this month.
- A hard length cap: 4 bullets max, 100-120 words total.
Pasting bullets instead of the raw deck is not just tidier — it is what keeps the whole report inside the model’s working window. As of June 2026, ChatGPT Plus runs GPT-5.5 Thinking at a 256K-token in-app context; the full 1M-token window is reserved for the $200 Pro plan. A 40-slide quarterly deck with embedded charts can exceed the Plus limit, at which point the model silently summarizes only the part it can still see. Twelve lines of bullets never hit that ceiling.
Which model to use (June 2026)
If your deck is short and you only paste bullets, the choice barely matters — pick whatever you already pay for. The model only becomes a factor when you upload the full file.
| Model | In-app context | Whole-deck upload | Best for this task |
|---|---|---|---|
| GPT-5.5 (ChatGPT Plus $20) | 256K (Thinking); 32K (Instant) | OK for short decks; full 1M only on $200 Pro | Bullet-paste workflow; crisp imperative prose |
| Claude Sonnet 4.6 (Pro $20) | 1M tokens | Up to 600 PDF pages per request | Long decks pasted whole; tone control |
| Gemini 3.1 Pro (Google AI Pro $19.99) | 1M tokens | Up to ~900-page PDFs in one prompt | Decks living in Google Slides / Workspace |
Sources: OpenAI GPT-5.5 model docs, Claude context windows, Gemini 3.1 Pro model card.
For a deeper side-by-side on which assistant fits which job, see ChatGPT vs Claude vs Gemini.
Copy-ready prompt
You are writing the exec summary for a 12-slide management report.
Per-slide bullets:
1. Revenue: 12.4M vs 12.0M plan; enterprise renewal pulled early
2. Margin: 56.2% vs 58% plan; input costs up
3. Sales OpEx: 2.8M vs 2.5M plan; reqs filled
4. Marketing: 0.9M vs 1.2M plan; Q3 push-out
5. Pipeline: 18.4M (+8% QoQ); enterprise heavy
6. Churn: 4.2% vs 4.0% plan; one large logo
7. Product: feature X shipped; adoption 22%
8. NPS: 38 (-2 vs Q2); pricing complaints up
9. Hiring: 4 of 6 open reqs closed
10. Market: competitor A raised price; we did not
11. Risk: supplier B contract up in Q4
12. Q4 outlook: revenue 13.2-13.6M; margin 54-55%
CEO will ask about: input costs, pipeline quality, NPS trend.
Write:
- 4 bullets, ranked. Each <= 25 words.
- Each bullet ends in a verb the CEO can act on (defer, hire, escalate, hold).
- No percentages they cannot see on the cover slide.
- Strip all "we should consider" language. Use direct language.On GPT-5.5 select Thinking, on Claude use Sonnet 4.6, on Gemini set thinking to Medium — ranking bullets by likely behavior change is a reasoning task, and the fast/instant modes tend to keep all 12 lines instead of cutting to 4.
What good output looks like
- Margin compression is the story, not revenue. Input costs are structural; Q4 lands at 54-55% unless we renegotiate supplier B. Escalate to procurement this week.
- Pipeline looks healthy ($18.4M) but enterprise-heavy. Single-deal risk is real — the top 5 logos are 38% of pipeline. Defer non-enterprise reps until coverage rebalances.
- NPS down 2 on pricing. Competitor A held; we held. We are not losing on product. Hold the pricing review for 4 more weeks of data.
- Hiring stalled at 4/6. Two open SDR reqs are the bottleneck for Q4 pipeline. Hire the SDR manager before quarter-end.
How to refine
- Six bullets came back — push back: “Pick 4. Drop the two weakest.”
- Bullets too long — add a hard limit: “Each bullet <= 20 words. Cut connective tissue.”
- The model hedges — strip “could”, “should consider”, “may want to”. Replace with imperatives.
- Wrong ranking — instruct it: “Rank by what is most likely to change behavior, not by financial size.”
Common mistakes
- Pasting the full slide deck as-is. The model produces a longer summary, not a shorter one.
- Leaving in narrative (“In Q3 we saw…”). Cover slides do not narrate; they decide.
- Including bullets the CEO already knows from yesterday’s 1:1. Strip those.
- Forgetting the “what to do” verb. A summary without verbs is a recap, not a brief.
FAQ
- What length actually fits one screen? Four bullets is the upper bound for a single phone screen. At five or more, the CEO reads only the first four — so cut the fifth yourself rather than letting it get skipped.
- Can I make this recurring? Yes. Save the prompt with the 4-bullet rule and the verb rule as a reusable “monthly mgmt brief” prompt, then swap in fresh per-slide bullets each cycle. See decision summary prompt templates.
- What if my report has no clear outlook slide? Add one before you summarize. The model cannot synthesize a forward view from backward-looking data — it will either hedge or fabricate a number, and neither belongs on a CEO cover slide.
- Should I upload the whole deck or just bullets? Just bullets, in almost every case. It keeps the report inside the context window on a $20 plan and forces you to do the editorial thinking the model cannot. Upload the full file only when the deck is too large to hand-extract and you are on a 1M-context model (Claude Sonnet 4.6 or Gemini 3.1 Pro).
- Will the numbers be accurate? The model copies figures you paste; it does not recompute them. It can still misattribute a number to the wrong line, so always check that each percentage in the summary matches the slide it claims to come from.
Related
- AI variance analysis
- AI KPI weekly report
- AI business driver breakdown
- AI sensitivity analysis
- Decision Summary Prompts: 15 Templates for Memos People Actually Read
Tags: #AI writing #Finance #Summary #Executive update #Business analysis