talk-script-clarity

Talk Script Clarity

Tailor the presentation script to an audience that hears it once and is done. A written text lets the reader go back and read it again, but a spoken text is over once it passes by. So even for the same content, repetition, metaphor, numbers, and sentence length are handled differently from written text.

These principles were organized by referring to the structure of Professor Choi Ye-jin's TED talk ("Why AI is incredibly smart and shockingly stupid", 2023). The original talk text is not included. The principles were reorganized by the worker on this repository, and are not the original speaker's official guidelines.

When to use

When the slides exist and you want to polish the speaker notes
A request to revise it to be "natural", "crystal clear", and "easy to understand by ear"
When checking whether the script flows before and after after changing the deck order
Final check before a defense, conference, or lab seminar rehearsal

When NOT to use

Creating the deck from scratch → academic-slides-iterative
Polishing paper paragraphs → paper-section-rewrite (the rules for written text are different. See "Differences from written text" below)
Fixing facts or numbers themselves. This skill changes only the way of speaking

7 principles

Principle 1. Plant a one-sentence theme at the beginning and the end

Write in one sentence the picture the entire presentation should leave behind. This sentence appears on the first slide (or in the first minute), and appears again in almost the same words on the conclusion slide and the final slide. On the conclusion slide, say this sentence separately and pause for a moment. For each section, also place a one-line summary, and check whether connecting the section summaries produces the presentation theme sentence.

Check: Place the theme sentence of the first slide and the theme sentence of the conclusion slide side by side and see whether they say the same thing.
Failure mode: Summarizing with a new expression heard for the first time in the conclusion. The audience does not know whether it is the same story.

Principle 2. Carry one big metaphor from start to finish

Choose one metaphor that carries the presentation's sense of the problem (e.g., "the trap of falling into a familiar solution"). This metaphor is ① planted early, ② named, ③ called back with the same word in each middle section, and ④ retrieved in the conclusion and the final slide. Do not replace the metaphor word with another word.

Deliverable: metaphor retrieval map — the section numbers where the metaphor appears and its role in that section (planting/naming/retrieval/closing).
Check: If empty stretches in the map exceed 1/3, create one more middle retrieval point.
If there is no metaphor, propose 2–3 candidates from the script's repeated words and sense of the problem, then choose one.

Principle 3. For each core concept, provide an example the audience can work through in their heads

First show an everyday example or analogy the audience can try out in their heads, then show the data. The order is example → what the model (or method) does → why it is so → what it means.

Target: 2–3 core claims of the presentation. Do not put them in every chapter.
Conditions for a good example: it can be brought to mind within 10 seconds without specialized knowledge. It shows only one key difference in the concept.
E.g., "Raising the temperature is like blurring the text on a menu board. Item 1 is still item 1. RSP shuffles the order of the menu board."

Principle 4. One number per claim

Numbers spoken aloud do not stick in memory. For each claim, say only the single most important number (at most one contrasting pair), and leave the rest on the slide.

Criterion: if the number of numbers spoken per slide is ≥ 8, that is a warning; if the numbers per slide exceed two contrasting pairs, cut some out.
How to cut: turn precise values into approximations ("close to 20,000 games," "more than five times," "twice as much"), move setting numbers (sample size, number of repetitions) to the slide, and when there are three or more comparison targets, say only the two extremes.
Safeguard: when changing a number to an approximation, always check that the direction and magnitude match the original value. Saying an approximation together with the exact value, such as "more than half, 57.8%," is allowed.

Principle 5. Unpack terms when they first appear, and use one name per concept

When saying an abbreviation or technical term for the first time, attach a plain-language explanation ("sparse autoencoder, SAE for short," "LoRA, which trains only a small additional weight").
Mathematical terms (logit gap, upper bound, rank correlation, tautology, etc.) should be turned into everyday language or defined once ("the score difference between the top candidate and the remaining candidates," "at most this much," "if it is 1, the order has not changed at all").
One name per concept. If the paper's term and the presentation's term differ, connect them only once ("the paper calls it injection, but today I'll call it attaching"). After that, do not change it.
Check: count the occurrences of the group of names referring to the same target (e.g., injection/attaching/inserting, branching/fork/offshoot) and see whether one name dominates.

Principle 6. Announce the number in advance, and build bridges between chapters

When saying several things, state the number first ("three questions," "two conditions").
The last sentence of a chapter becomes the question of the next chapter. The first sentence of the next chapter starts by picking up the last words of the previous chapter.
Do not let the same connective ("however," "so") appear back-to-back at chapter boundaries.
After changing the order or merging chapters, always read the whole thing through again. Seams usually break right before and after the place you changed.

Principle 7. Short sentences, and the conclusion chapter should be the shortest

One thought per sentence. Sentences longer than 50 Korean characters are candidates for splitting.
Proportion of short sentences (≤ 50 characters) per chapter ≥ 60%. However, if four or more short sentences in a row appear without connectives, it sounds like bullet points, so combine cause and effect into one sentence.
The conclusion chapter (and the final chapter) should have the shortest sentences in the presentation. Mark "(pause)" before the topic sentence.

Work order

Step 0  입력 정리      대본을 장 단위로 나눈다. 형식: [번호 · 제목 · 초] 본문
Step 1  주제·은유      한 문장 주제와 큰 은유를 뽑는다(없으면 후보 제안 후 선택)
Step 2  은유 회수 지도  은유가 등장하는 장과 역할 표. 빈 구간 표시
Step 3  측정(전)        scripts/clarity_metrics.py 실행. 숫자·문장 길이·용어·이름 묶음·은유 등장
Step 4  장별 점검·수정  7원칙 체크리스트(references/checklist.md)로 장마다 고친다
Step 5  이음새 통독     전체를 처음부터 끝까지 읽으며 장 경계만 본다(원칙 6)
Step 6  측정(후)·보고   같은 스크립트로 다시 재고 전후 표를 낸다. 장 시간 변화도 적는다

Step 0 — Organize input

If the script is in the HTML deck's data-speaker-notes, read it as is. If it is Docs or Markdown, [번호 · 제목 · 초] split it by heading units.
Also look at the slide screen text. If the script and the screen state different numbers, align to the screen.

Step 3 — Measurement

python scripts/clarity_metrics.py deck/index.html --metaphor 함정 \
    --aliases "주입,붙이기,끼워 넣" --aliases "분기,갈림길,갈래"
python scripts/clarity_metrics.py script.md --metaphor 함정      # [NN · 제목 · 초] 형식

Output: character count per chapter, average sentence length, short-sentence ratio, number count (≥ 8 warning), terminology list, name-cluster count, list of chapters where metaphors appear, work-history words, overall summary. Flags are judgment material. There are cases where something trips a rule but is not a problem, such as ranking names ("menu 1") or repeated experiment descriptions ("even if you run it again").

Step 4 — Per-chapter revision rules

Do not change facts or figures. What changes is only the way of speaking. For numbers changed into approximations, check whether the direction and magnitude match the original value.
Check whether words pointing to figures ("left", "top", "red line") match the actual slide layout. If you changed the slide, review the direction words in the script too.
Chapter time notation ([45초]) is revised to match the change in character count (Korean spoken language is about 5–5.5 characters/second).
Replace work-progress expressions with content names ("review response update" → "5 seed average", "coming out of the review" → "to answer this question"). The audience does not know the work history.

Step 6 — Report format

| 항목 | 전 | 후 |
|---|---|---|
| 말하는 숫자 합계 | 155 | 132 |
| 짧은 문장 비율 | 77% | 82% |
| 풀지 않은 전문 용어 | 로짓 간격, 항진식, … | 0 |
| 은유 등장 장 | 0 | 11 (1, 10, 11, 13, 23, …, 50) |
| 장 머리말 시간 합계 | 37분 20초 | 36분 50초 |

Write per-chapter changes as "chapter number · principle fixed · one-line summary".

Differences from reading text (when used together with paper-section-rewrite)

This skill follows the spirit of paper-section-rewrite's Levels 1–4 (content → between paragraphs → within paragraphs → sentences). Remaining figures, sentence connections, the non-expert gate, and history independence apply as is. However, because it is text heard by ear, the following four are applied in the opposite way .

Point	Reading text (paper-section-rewrite)	Listening text (this skill)
Repetition	Saying the same thing three times is muddled	Topic sentences and metaphors are deliberately repeated at the beginning, middle, and end
Concept words	If 5 or more times, rotation is recommended	Metaphor words and defined core terms are never rotated
Style	Mixing in colloquial style is register ⚠	Colloquial style is the default. "~yo", "~jyo" are allowed
Count signals	"First, second" is mechanical parallelism ⚠	Announcing the count is an essential device
Negative lead	Zero negative first sentences in a paragraph	A negative hook that flips expectations is allowed (once per chapter)

In addition, humanize-korean's rules banning translationese and copula substitutes apply as is.

Skills used together

academic-slides-iterative — After building the deck, check the companion talking points' "slide-by-slide script" with this skill
paper-section-rewrite — When writing the same content as a paper. You only need to remember the differences in the table above
humanize-korean — Organizing Korean expressions that sound AI-generated

Supporting files

references/checklist.md — One-page checklist per chapter of the 7 principles
references/examples-rsp.md — Before-and-after examples revised from an actual lab seminar script (RSP) and a metaphor callback map
scripts/clarity_metrics.py — Quantitative measurement per chapter