slug: core-message-extraction view: skill layer: slide agent: architect companion: corpus/storymakers/frameworks/slide/core-message-extraction.md audience: llm

Core message extraction — operational reference

Compact reference for an LLM extracting the one-sentence core message from raw slide content. The showcase (core-message-extraction.md) is the human-facing explanation; this is the chuleta.

One sentence. Look at the raw content (chart, paragraph, analysis) and refuse to ship the slide until you can write one declarative sentence with a verb — the sentence that earns the headline.

Use this when

  • The slide carries an argument or finding, not pure reference.
  • The current title is a topic, not a claim (no verb, no stake).
  • A chart or analysis sits alone without a sentence defending it.
  • A reader could draw more than one conclusion from the slide.

Use something else when

If… Then…
Slide is appendix / sources / disclaimer No core message — let it be reference.
Deck is exploratory, answer not found Mark slide "open question"; don't fake a core.
Multiple equally-important claims Split into multiple slides, one core each.
Summarising a deck, not a slide Use governing-thought / big-idea-formula.
The artefact is the title, not the process Use action-titles directly.

Decision tree

Q1. Does the slide carry a claim (vs pure reference)?
    NO  → no core message needed. STOP.
    YES → Q2

Q2. Have you found the core — one declarative sentence with a verb?
    NO  → analysis incomplete or sentence hiding. Run the recipe.
    YES → Q3

Q3. Does it pass the so-what test?
    NO  → climb. Add the implication clause. Re-test.
    YES → Q4

Q4. Could the audience act on it (even if "keep going")?
    NO  → trivia, not a core message. Climb or cut.
    YES → Q5

Q5. Is this the highest defensible claim from the evidence?
    NO  → climb until evidence pushes back.
    YES → ship as the action title.

If the answer to any question is NO, return to Q2 and re-extract.

Recipe — the extraction loop (~15 min per slide)

  1. Read the raw content. Chart, paragraph, table, interview notes. Do not look at the existing title. (2 min)
  2. Draft a candidate sentence. Subject, verb, defensible claim. Don't optimise yet — commit. (3 min)
  3. Run the four tests, in order: (5 min)
    • Verb test. Is the verb doing work? ("shows, is, varies" are weak; "reverses, destroys, doubles" are strong.)
    • So-what test. Add the implication clause. If it's already there, check it earns its place.
    • Action test. Complete "…so we should…". If you can't, climb.
    • Highest-claim test. Is there a stronger sentence the same evidence would support? If yes, write it. Repeat.
  4. Compress. Cut hedges, parentheticals, dates the audience already knows. Target ≤20 words. (2 min)
  5. Promote to action title. The extracted sentence becomes the slide's title. The chart, table, or paragraph below it now has one job: defend the sentence. (2 min)
  6. Falsify. Could a senior reader argue with it? Good — that means it's a claim. If they can't, it's a topic; restart at Q2. (1 min)

Headline language (verbatim bank)

  • "[Subject] [strong verb] [magnitude]; [implication clause]."
  • "[Finding] — [so-what], so we should [move]."
  • "[X] is [Y]; the [structural shift] this implies is [Z]."
  • "Not [obvious-conclusion]: [sharper-conclusion-the-evidence-supports]."

Anti-patterns (reject on sight)

Pattern Why it's wrong
"Revenue performance", "Customer satisfaction" Topics, not claims. No verb, no stake.
"X grew, Y declined, Z launched" Multi-thesis stapler. Pick one.
"Some signs may suggest a possible softening" Hedge stack. Drop the adverbs or cut the slide.
"Q3 revenue fell 4%" (as the title) Chart caption. Add the so-what.
50-word sentence with three parentheticals Defend in supports, not in the headline. Compress.
"This slide explores…" / "…provides an overview" Meta-title. Cut. Say what the slide concludes.
"In essence, X is X" Tautology dressed as insight. Climb to the actual finding.
Restating Q1's draft as the final answer Stopped at description. Climb to implication.

Output checklist

A slide is shippable when:

  • Title is one declarative sentence with a verb.
  • The sentence carries its own so-what on the page.
  • Reader could complete "…so we should…" without help.
  • Claim is the highest defensible — climbed until evidence pushed back.
  • Sentence is ≤20 words.
  • Chart / paragraph below now functions as evidence for the title.
  • A reasonable reader could disagree — i.e. it's a claim, not a topic.

Canonical signals (for matchers)

  • Keywords: bottom line, key insight, the takeaway, in essence, in short, core finding, the point is, nut graf, tl;dr, commander's intent, find the core, one-liner.
  • Component kinds: title (action-title style), callout (the extracted sentence highlighted), subtitle (deck/section apex).
  • Slide-level shape: a single declarative sentence at the top, with a chart / table / paragraph below that exists to defend it.

References

Heath & Heath, Made to Stick (2007), ch. 1; Blundell, Art and Craft of Feature Writing (Plume, 1988); Roy Peter Clark, The Nut Graf and Breaking News (Poynter, 2003); U.S. Army, FM 3-0 Operations (2022); Kahneman, Thinking, Fast and Slow (2011), ch. 7 (WYSIATI); Minto, The Pyramid Principle (1973).