methodsguide

How to Build a Presentation with AI Without Losing the Argument

An argument-first sequence that separates evidence, narrative, and slide production.

Source fragments and an argument strip lead to a sequence of paper slide layouts.
Last reviewed 2026-10-13

Updated: This article was revised on .

A presentation can look finished while its argument is still missing. Generating slide titles too early encourages the team to decorate a sequence before deciding what the audience should understand, believe, or do.

Use AI after the reasoning is visible. The sequence is argument → evidence → structure → slides → visual support. Moving backward is allowed; skipping directly to slides is the failure to avoid.

For a detailed handoff, keep the source-to-slide chain separate.

What a bounded field check showed

This method was checked in one small editorial exercise on 6 August 2026. A synthetic packet held one decision, six labelled observations, one deliberate contradiction, an output specification, and the same five-part rubric in both conditions. One condition composed directly; the other inventoried notes, mapped claims, challenged the contradiction, and then composed.

The staged condition retained all six source IDs and the contradiction. The direct condition merged two observations. The reviewer scored traceability, specificity, contradiction handling, usability, and repair effort from 1–5; the result was 5/4/5/4/4 for the staged pass and 2/3/2/4/2 for the direct pass. This is a dated method observation, not a model benchmark: one reviewer created the fixtures, knew the conditions, and no external model or private account was tested. The useful lesson is narrower: make evidence and contradiction checks visible before slide composition.

Write the audience change

Before opening a tool, complete three sentences:

  • The audience currently believes or does…
  • After the presentation, they should understand…
  • The decision or action we are asking for is…

If the requested change is vague, more slides will not fix it. Name the strongest reasonable objection too; it will shape the evidence you need.

Build an evidence table

Create one row per consequential claim. Include source, date, exact support, limitation, and the audience question it answers. Separate evidence from interpretation. A chart is not evidence by itself: its source, denominator, time period, and relevance must remain available to the reviewer.

Draft the argument as sentences

Write a short chain before slide titles:

  1. Situation the audience recognizes.
  2. Tension or problem that changes the stakes.
  3. Claim supported by the evidence.
  4. Alternatives considered.
  5. Recommendation and consequence.

Ask AI to challenge missing warrants or reorder the chain, not to invent support. Keep rejected structures; comparing them can reveal which assumption is doing too much work.

Convert one claim into one slide job

A slide job is a sentence such as “Show that delays concentrate at the handoff, not throughout production.” It is not “Operations overview.” From each job, derive a headline that states the point, the minimum evidence required, and the speaker’s transition.

For each proposed slide, fill this small record before asking for visual ideas:

Audience question:
Claim the slide must make:
Source and exact supporting passage:
Limitation or alternative explanation:
Slide job in one sentence:
Visual needed to make the claim clear, if any:

If the source line is empty, treat the slide as a question to investigate or an opinion to label, not a factual claim ready for design.

Only then decide whether the best support is a chart, excerpt, diagram, table, or no visual at all. Decorative imagery should not compete with evidence. Every element must either advance the claim, provide orientation, or be removed.

Illustrative before and after

Before: “Market opportunity” followed by four uncited statistics and a generated illustration.

Process: The editor checks the sources, discovers that two figures describe different geographies, and narrows the claim to the one decision the available evidence supports.

After: “The reachable segment is smaller—and better matched to our channel” with one sourced comparison, a visible assumption, and a note explaining what would change the estimate.

This example demonstrates structural improvement; it is not a report of measured presentation performance.

What happened when rough notes became slides

In a separate test, we gave the same AI assistant invented, messy notes about whether a library should trial Saturday evening opening. The notes included possible access benefits, no evening footfall data, uncertain staffing, dim lighting, and a budget meeting that might delay a pilot. The first six-slide outline kept the central tension: later hours could help some people, but demand was unproven and visits might simply move from the afternoon.

One focused revision made the committee action clearer: prepare a reversible trial rather than announce permanent hours. It also pulled an unconfirmed printer-lease date from the notes into the budget slide. That detail had no demonstrated bearing on evening opening. The revision was mixed: a sharper decision, plus a newly promoted distraction. Neither outline proves a real library should change its hours.

For each slide, use a short record that keeps the source visible:

Slide claim: [one sentence]
Supporting note: [exact note or source]
Uncertainty or alternative explanation: [what remains open]
Decision this slide helps the audience make: [decision]

After revising, check whether background details have gained weight they did not earn. A clearer sequence is only useful if its claims still match the notes.

A critique prompt

Review this argument map as a skeptical audience member. Do not write slides. For each claim, identify the evidence supplied, the missing warrant, the strongest alternative explanation, and the decision it supports. Mark unsupported claims rather than completing them. Return a table keyed to claim IDs.

Use the output as questions for the author. Do not accept an objection or correction without checking it against the source set.

Try, check, and save

Try: Take an existing deck and write the job of each slide in the margin. Remove slide language from the job statement.

Check: Can the audience action be stated in one sentence? Does every consequential headline have adequate support? Does the sequence answer the strongest objection before the recommendation?

Save: Use the presentation argument planner to keep audience change, claims, evidence, and slide jobs together.

Limits

AI cannot decide which evidence your organization is permitted to use, whether a recommendation is politically or ethically sound, or whether the presenter truly believes the argument. Generated charts and citations require direct verification. For high-stakes decisions, involve the relevant domain and accessibility reviewers.

The deck is ready for design when the argument can survive without design. Visual craft should sharpen reasoning, not disguise its absence.

Practical resources

Evidence ledger

Sources

  1. 01Prompt engineering overview
  2. 02Prompt engineering best practices

Links are descriptive and separated from editorial conclusions. Product behavior may change after the review date.

Written by

Raúl Mermans

Raúl Mermans creates and edits Token & Taste, documenting practical AI methods and reviewing the evidence behind its guidance.

About the author →

Next methods

Continue the work

Or browse Latest