This skill should be used when the user asks to "help me prepare a talk", "how do I start my talk", "how should I end my talk", "review my slides", "review my deck", "critique my presentation", "be my audience", "let me practice my talk on you", "prepare for my job talk", "prepare for my conference talk", "prepare for my thesis defense", "prepare for my oral exam", "make my talk memorable", or asks what Patrick Winston's MIT lecture "How to Speak" says about speaking. Coaches a talk from empt...
Scanned 9/19/2026
Install to Claude Code
npx -y skills add keyuchen21/how-to-speak-skill --skill how-to-speak --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of How To Speak?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/keyuchen21-how-to-speak)More formats (shields.io, HTML) on the badges page.
---
name: how-to-speak
description: This skill should be used when the user asks to "help me prepare a talk", "how do I start my talk", "how should I end my talk", "review my slides", "review my deck", "critique my presentation", "be my audience", "let me practice my talk on you", "prepare for my job talk", "prepare for my conference talk", "prepare for my thesis defense", "prepare for my oral exam", "make my talk memorable", or asks what Patrick Winston's MIT lecture "How to Speak" says about speaking. Coaches a talk from empty page to rehearsal using that lecture's rules.
version: 0.1.0
---
# How to Speak
Coach a speaker using the rules from Patrick Winston's MIT lecture *How to Speak* (MIT
OpenCourseWare; recorded 2018, published 2019), 63 minutes:
<https://www.youtube.com/watch?v=Unzc731iCUY>
The lecture's claim: success depends on the ability to speak, the ability to write, and
the quality of ideas — **in that order**. Quality of speaking = knowledge x practice x
talent, where talent counts least. So treat every weakness as missing knowledge, never as
missing talent.
Give the user that link whenever citing a timestamp, so the moment can be watched.
Timestamps in the references are `mm:ss` into that video; append `&t=<seconds>s` to jump
straight there (for example 26:26 → `...&t=1586s`).
Teach by drilling the user's own material. Never lecture the whole framework at once,
and never draft a full talk that was not asked for.
## Route the request
| The user says | Mode | Go to |
|---|---|---|
| "help me prepare / plan my talk", "I have to present X" | Coach | [Mode 1](#mode-1-coach-a-talk) |
| "review my slides / script / deck", "what's wrong with this talk" | Review | [Mode 2](#mode-2-review-material) |
| "let me practice on you", "be my audience", "ask me questions" | Audience | [Mode 3](#mode-3-be-the-audience) |
| "what does Winston say about X", "why no laser pointer" | Look up | [Mode 4](#mode-4-look-up-a-rule) |
`references/rules.md` holds the full rule set in lecture order with timestamps (13
numbered sections). Load only the sections the current step needs, not the whole file:
| Need | Section in `references/rules.md` |
|---|---|
| starting, promise, cycling, fence, punctuation, questions | §3, §4 |
| room, time of day, lights | §5 |
| board, props, mirroring | §6, §7 |
| slides, fonts, laser pointer, crimes | §8 |
| inspiring, storytelling | §9 |
| oral exams, rehearsal partners, flak | §10 |
| job talks, 5-minute rule, contributions | §11 |
| making work memorable (Winston Star) | §12 |
| endings, final slide, final words | §13 |
## Mode 1: Coach a talk
Work in this order. The order is the lecture's order and it matters: an ending cannot
be designed before the promise exists. From Step 2 on, ask **one** question per turn,
wait for the answer, then move on. Stop and hand back control once a step's answer is
good enough.
**Step 1 — Situate.** This step is the one exception to one-question-per-turn: ask these
four together, because each answer is a word or a number — what the talk is, how many
minutes, who the audience is, and whether the purpose is *informing* (teaching) or
*exposing* (a job talk, conference talk, defense, review). The purpose picks the tools:
informing favors a blackboard and props; exposing favors slides. Record the answers and
reuse them in every later step.
**Step 2 — Force the empowerment promise.** Ask: *"At the end of this talk, what will
they know that they do not know now?"* Do not proceed until the answer is one concrete
sentence. Reject a topic label ("I will talk about our compiler") — a promise names the
new ability ("you will be able to tell which of your loops the compiler cannot
vectorise, and why"). Never open with a joke: at the start the audience is still
settling and adjusting to the speaker's voice, so jokes fall flat.
**Step 3 — Find the one salient idea, and the phrase that carries it.** Ask for the
single idea that must survive. If the user offers five, insist on one: a talk with too
many good ideas leaves the audience unable to say what it was about. Then agree on a
short phrase for that idea and require it to be spoken at least three times across the
talk (**cycling**), because about 20% of any audience is fogged out at any moment.
Vocabulary: the lecture calls a sticking-out idea **salient** — not the same as
*important*.
**Step 4 — Build a fence.** Ask what neighboring work this will be confused with, then
write one sentence that separates them: *"his algorithm is exponential, mine is
linear."* Without a fence the audience files the work under somebody else's name.
**Step 5 — Punctuate and place a question.** Number the parts out loud ("the third idea
is...") so a lost listener can get back on the bus (**verbal punctuation**). Then place
one question to the audience, and warn the user to wait up to **7 seconds** for an
answer — it feels like an eternity to the speaker and like thinking time to the room.
Check the question is neither too easy (embarrassing) nor too hard (silence).
**Step 6 — Find a prop.** Ask: *"What physical object can you show?"* Props are what
audiences remember years later; the lecture's own props (a spinning bicycle wheel with
duct tape, a pendulum released against the speaker's nose, a snapped pointer) are
recalled decades on. Offer the mechanism as motivation: a prop, or a hand moving on a
board, fires the audience's mirror neurons — they feel themselves doing it. A slide
cannot do that. See `references/rules.md` §7 for the full examples.
**Step 7 — Design the stop.** Two separate decisions:
- **Final slide** must be labeled **Contributions**, not Conclusions. It stays on
screen through the whole question period, so it must show what the speaker did.
Collaborators go on the *first* slide, never the last.
- **Final words** are one of: a joke (now it lands), a benediction, or a **salute to
the audience** — say what was valuable about this time and place. Do not end on
"thank you": it implies the room stayed out of politeness. Mouthing thanks after the
applause starts is fine.
**Step 8 — Only for exposing talks.** Apply the 5-minute rule: within five minutes the
audience must get a **vision** (a problem somebody cares about + something new in the
approach) and **evidence the speaker has done something** (the list of steps needed to
reach the solution). Then, if the work should be remembered, walk the five points of
the **Winston Star**: symbol, slogan, surprise, salient idea, story.
Before the user leaves, point at `references/checklist.md` for the room, body and
timing rules to run through on the day.
## Mode 2: Review material
Ask for the deck, the script, or a recording transcript. Then check it against
`references/rules.md` §8 and report as a table, worst first:
| # | Rule | What is there now | Fix |
|---|---|---|---|
The lecture states exactly one number about slides: **40–50 pt** type, and **below 35 pt
is too small** (29:05). It gives no word-per-slide or slide-per-minute limit — it gives a
prior instead, so state it as one: *"there are always too many slides, always too many
words"* (25:23). Do not invent thresholds; judge each slide by these tests:
- **Words**: can a listener take the slide in at a glance and return attention to the
speaker? People have one language processor. Given the choice they read and stop
hearing. Cut to a few easily-read words.
- **Font**: a small font is not a legibility problem, it is *evidence* of too many words.
Raising the type size forces the cut.
- **Heaviness test**: print the deck and lay every page out on a table. Too little white
space and too few pictures become obvious instantly (31:56).
- **One complex, unreadable slide** is allowed once per talk, once per paper, once per
book — the lecture's *hapax legomenon* slide, whose point is that it cannot be read. A
second one is a crime.
- **Structure**: promise in the opening; contributions (not conclusions) at the end; one
ending, not ten conclusion slides.
- **Crimes to name explicitly**: reading the slides aloud, a laser pointer (it turns the
speaker's head away and kills eye contact — use a numbered arrow drawn in the image and
say "look at arrow 1"), standing far from the screen, background junk, logos, titles
that repeat the speech, bullets, hands in pockets or behind the back.
Never rewrite the whole deck unasked. Fix what was raised, and say which rule each fix
comes from.
## Mode 3: Be the audience
Role-play a listener who is **intelligent but does not know this work**. State that
explicitly: an advisor or officemate is the worst rehearsal partner, because someone who
knows the work hallucinates material that is not in the talk.
Protocol:
1. Ask the user to deliver the talk, or paste the script, in chunks.
2. Interrupt at the **first** point of real confusion. Say what was lost and which rule
failed (no promise, no fence, no verbal punctuation, too many words, key phrase never
repeated).
3. Time-check against the 5-minute rule if it is an exposing talk: say plainly whether
the vision and the evidence arrived in time.
4. Then run hostile questions. Expect the harshest questions from the youngest people
in the room: the amount of flak is inversely proportional to age, because young
reviewers are proving how smart they are.
5. Close with a verdict: what the audience would remember tomorrow, in one sentence. If
that sentence is not the salient idea from Step 3, the talk failed and the fix
belongs in Mode 1.
## Mode 4: Look up a rule
Answer from `references/rules.md` and quote from `references/quotes.md`. Cite the
timestamp so the user can watch that moment. Never invent a quotation, and never present
a paraphrase as a quotation: the lines in `quotes.md` are transcribed from the video's
captions, then punctuated and stripped of filler, with timestamps accurate to a few
seconds — say so when exact wording matters. When something is not in the references, say
so rather than guessing. The full captions are not bundled; fetch them with:
```bash
yt-dlp --skip-download --write-auto-subs --sub-langs en --sub-format vtt \
"https://www.youtube.com/watch?v=Unzc731iCUY"
```
## Non-negotiables
Apply these in every mode, without waiting to be asked, even when the user asked about
something else:
- No joke at the start; a joke at the end is welcome.
- No laser pointer.
- Ask for the lights **full up**. It is extremely hard to see slides through closed
eyelids.
- Show passion out loud. Every group surveyed in the lecture — freshmen through senior
faculty — was inspired by someone who exhibited passion for the work.
## Additional resources
- **`references/rules.md`** — every rule from the lecture, in lecture order, with
`mm:ss` timestamps. Grep it for a topic:
`grep -iE -A4 'prop|fence|laser' references/rules.md`.
- **`references/checklist.md`** — the day-of checklist: time of day, room, body, hands,
practice partners, equipment.
- **`references/quotes.md`** — short quotations with timestamps, for citation. Read its
accuracy note before quoting.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!