methodsworkflow

From Research to Slides Without Letting AI Flatten the Argument

A five-stage workflow that preserves evidence, causality, and narrative before AI-assisted slide production.

Five connected paper cards show the path from source to claim, argument, narrative, and slide.
Keep the source, claim, argument, narrative, and slide distinct until each handoff is reviewed. — Original Token & Taste diagram

Updated: This article was revised on .

AI can turn a source packet into a plausible deck very quickly. The failure is not that the deck is ugly. It is that a source becomes a sentence, a sentence becomes a slide title, and the missing reasoning disappears between them.

In one small test, the five-slide outline kept the key counts and the missing follow-up caveat visible. Its recommendation to repeat the workshop was a judgment drawn from a fictional packet, not a measured outcome. To check an AI presentation, map every headline and number back to a source ID, and mark recommendations as recommendations.

A source-to-slide check we ran

On 15 October 2026 we supplied gpt-5.6-sol with four short synthetic records about a fictional library workshop and asked for five slides with source IDs beside facts. The retained packet and raw outline show the exact input and output. We reviewed each slide against the records.

Slide claim Source check Judgment
Six of eight registrants attended a 45-minute session. S1 says exactly that. Supported.
Four respondents identified one decision AI should not make. S2 says four of five returned cards did; the slide states the denominator. Supported.
Three of five respondents completed the checklist. S2 supports the count; it does not show home use. Supported with scope.
A paper backup preserved the session after Wi-Fi failed. S3 records the fallback and completed activity. “Preserved” is our interpretation. Inference, label it.
Repeat as a better-measured pilot. S3 supports planning a backup; S4 says there is no later outcome evidence. Recommendation, not a finding.

What we found: this outline did not invent a statistic or claim improved real-world paperwork handling. It did turn the recorded fallback into the interpretive headline “preserved the session,” which a reviewer should label as an inference. One synthetic run cannot establish that the model reliably preserves caveats.

Before a deck goes out, use this short review: underline every number and causal verb; find its source ID; check denominator, time frame, and missing evidence; label interpretations and recommendations; then remove any headline that outruns the packet. A copyable source-to-slide checklist keeps that pass beside the deck.

Keep five objects separate: SOURCE → CLAIM → ARGUMENT → NARRATIVE → SLIDE. Each answers a different question and deserves a different review.

The five handoffs

Source: record the passage, date, author, scope, and limitation. Claim: state only what the source can support. Argument: explain why that claim matters to the decision and what alternative explanation remains. Narrative: order the claims so the audience can follow the change in stakes. Slide: give one visual or verbal job to one moment in the narrative.

AI can help inspect gaps at every stage. It should not turn an absent warrant into confident connective language.

A concrete transformation

Imagine an internal report says: “In interviews with twelve support leads, handoff delays were mentioned more often after an escalation than after a routine request.”

  • Source: the dated interview memo; small sample, self-reported observations.
  • Claim: escalation handoffs were a more frequent concern in this interview set than routine handoffs.
  • Argument: improving routine response time alone may not address the delay the team is hearing about.
  • Narrative: first show why the broad service-time story is incomplete; then focus attention on the escalation moment.
  • Slide: “The reported delay is concentrated at escalation handoff.” Use a labelled excerpt count, the sample limitation, and a transition to the proposed review.

The source does not prove the whole service system is slow. The slide should not say it does. That restraint is the argument becoming more useful, not less persuasive.

Use AI as a skeptical assistant

At the claim stage, ask it to list unsupported modifiers and missing definitions. At the argument stage, ask for the strongest alternative explanation. At the narrative stage, ask it to identify the point where the audience has to make an inferential leap. At the slide stage, ask it to remove anything that does not serve the slide job.

Do not ask for “a compelling deck” while the source table is still unresolved. That instruction rewards rhetorical continuity over evidence discipline.

A short review pass

Before generating slides, check that every consequential headline can point back to a claim ID, every claim can point back to a source, and every source limitation survives the handoff where it matters. The presentation argument planner can keep those fields together. For a worked argument-first sequence and a bounded slide test, continue with how to build a presentation with AI.

This is a practical framework, not a report of an experiment. It does not claim that a particular model, prompt, or slide tool preserves arguments reliably. It gives a human reviewer the places where flattening usually begins.

Practical resources

Evidence ledger

Sources

  1. 01How to Build a Presentation with AI Without Losing the Argument
  2. 02Source-to-slide test packet

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