Conference Talk Craft

🗒️ Description

A researched playbook for running a 25-minute technical conference talk, assembled from 27 techniques across narrative, slides, delivery, rehearsal, Q&A and stage operations. Each technique was researched against its primary sources and its evidence graded, which turned out to matter more than expected: a third of the standard advice in this field is sound practice resting on a debunked justification.

The material is cut for a practitioner who speaks occasionally because they built something worth explaining — not a professional speaker. It assumes Polish delivery to a technical audience, markdown-based slides (Marp), and roughly 25 minutes plus 5 minutes of questions.

This is wave 1 of the research. A second wave covering slide design, code on slides, narrative structure and accessibility is outstanding. The full field-by-field data and the source list live in the workspace under _raw/research-workspaces/wystapienia-publiczne-prelekcje/.

🧩 The techniques, in the order they happen

Before you arrive

Technique	What it is
Word-for-word scripted opening	The first 60–90 seconds written as prose and rehearsed, not improvised. Memorise the seams — first sentence, handover sentence — and stay free in the middle, so there is no single point of failure.
Cold open patterns	Contrarian claim, opening with a failure, a question to the room. In a saturated room the strongest hook is the one least like the talks before it, not the most dramatic one.
Framing for a saturated audience	What to delete when earlier speakers already covered your subject. Repeating the “what is X” introduction burns the slot.
Modular time blocking	Semi-independent blocks, each with its own open and close, so cutting one does not break the talk.
Expansion joints	Designed cut points, so a shortened slot gets cut rather than rushed. The named antipattern for the alternative is Shortchanged.
Timed dry run	One full run aloud with a stopwatch, targeting roughly 85% of the slot.
Rehearsal ladder + declared plan	Silent read → aloud → standing → to camera → to one person → dress rehearsal, with the number and type of runs committed to in advance.
Modular drills	Target, measure and correct one narrow seam — the first 30 seconds, one block handover — instead of only running the whole thing.
LLM as red team	Generate the hardest questions with the asker’s motive and a 60-second ideal answer. Useful purely as input; the model cannot judge whether your answer is good.
Red-team question bank	Not the questions that recur, but the 15–20 that hurt. Four lines each: the issue in your words, a 40-second answer, the likely follow-up, the counter.
Demo strategy	Live, recorded, or narrated over a recording — chosen deliberately. The antipattern is Dead Demo: a demo filling time you have no content for.
Demo environment hygiene	Terminal and editor fonts at 18–22pt on a hall projector, notifications off at OS level, dedicated browser profile, predictive shell autocomplete disabled.
On-screen data leak	A different class of risk from a failed demo, because the recording is permanent. Scan with gitleaks, purge the identity surface, photograph every screen the audience will see and review it corner by corner.
Marp export constraints	Fragments exist only in HTML and vanish silently in PDF. A <video> slide exports as an empty player unless it carries poster=.
No-slides, no-internet plan	The narrative version of the talk on one card, plus a whiteboard version of the core diagram, plus a stated rule for how long you troubleshoot before abandoning the projector.
Physiology on the day	The variable that moves is hours spent talking in loud corridors beforehand, not what you drink backstage. Vocal-fold rehydration is measured in hours.

First 90 seconds

• Question policy announced immediately — questions at the end, on a signal, or in the hallway. Protects both the opening and the time budget.
• Audience interaction — a show of hands, counted, as the cheapest state change available. One device can do attention, retrieval, measurement and permission, but a question trying to do all four does none.
• Seeding the first question — a prepared opener for the case where nobody moves.

Middle blocks

• Energy in a bad slot — state changes every 10–20 minutes; props and prizes.
• Reading the room — diagnosis routed through pre-talk data and one counted probe, not through faces.
• Live narration over a recorded demo — you play the recording and talk to it live. Keeps control of the clock without losing the sense that something is happening.
• On-stage time control — cumulative checkpoints in the presenter notes, a stated tolerance band, and cuts executed at block boundaries rather than immediately.

Last 90 seconds and Q&A

• Answer protocol — stop moving, let them finish, paraphrase into the microphone, answer one thing in 30–60 seconds, close by moving rather than by asking.
• Repeating the question — for the microphone and the recording, since audience questions rarely reach the audio track. But never repeat a hostile framing verbatim.
• Hostile questions, false premises, “I don’t know” — hold the room without arguing.
• Closing after Q&A, not on it — 45–60 seconds written in advance so the last sentence of the recording is yours.

⚖️ What the evidence actually supports

This is the part that changed how I read every other speaking source. The practices below are all worth keeping. The reasons usually given for them are not.

Sound practice, wrong justification

Practice	The usual reason	What the research says
Change state every 10–20 min	”Attention drops after ten minutes”	Not supported (Wilson & Korn 2007). Bunce et al. 2010 measured lapses of under a minute. The same paper is the honest anchor: clicker questions and demonstrations lowered attention decline, and the benefit carried into the next segment.
Voice and body work	Mehrabian’s 7-38-55 ratio	That study concerned contradictory signals about emotion and liking, not informational delivery. Mehrabian asked people to stop extrapolating it.
Managing pre-talk nerves	Power posing	Author retracted; eleven replications found no behavioural effect. The evidenced alternative is arousal reappraisal — naming the activation as excitement (Brooks 2014, JEP: General).
Rehearse to 85% of the slot	The number itself	Folklore; published targets scatter across 80–90%. What is evidenced is the bias it patches: the planning fallacy (Buehler, Griffin & Ross 1994). “One hour of rehearsal per minute of talk” is likewise unsourced.
Paraphrase the question	”It makes the asker feel heard”	Weger, Castle & Emmett 2010 (n=180): paraphrasing raised the listener’s social attractiveness but produced no increase in felt understanding. The real reasons are mechanical — the microphone, the recording, thinking time, and ownership of the frame.
Close after Q&A	Recency effect, peak-end rule	Recency is precisely what vanishes under a filled delay (Glanzer & Cunitz 1966), which is a conference attendee’s condition. Peak-end derives from 60-second cold-pressor trials. The load-bearing argument is simply that the recording ends where it ends.
Moderate caffeine before speaking	”It dries the voice”	A 2021 systematic review in Journal of Voice found no voice measure adversely affected. The defensible reason is tremor stacked on adrenaline.
Avoid alcohol the night before	”It dries the vocal folds”	The mechanism is sleep architecture — fragmented second-half sleep and reduced REM (Ebrahim 2013).
The post-lunch slot is hard	”Blood diverted to digestion”	The early-afternoon dip is bi-circadian and appears without a meal. The difficulty is real; the folk explanation is not.
Drill one block in isolation	”Part-task practice is efficient”	Wightman & Lintern 1985 found part-task training less efficient in most cases, winning mainly for complex, cleanly divisible tasks and mostly under backward chaining.

Also debunked and worth not repeating: milk causes mucus, menthol lozenges soothe the voice.

Reverse myths — the fear is the error

• “I don’t know” costs credibility. It does the opposite. Mushkat & Mayo find it raises perceived trustworthiness against directive advice, explained “I don’t know” beats bare “I don’t know” on competence, and people systematically expect a larger penalty than materialises.
• Seeding the first question is a rescue for a dead Q&A. Carter et al. 2018, PLOS ONE — 247 seminars, 35 institutions, 10 countries. Mean six questions per seminar; a male first questioner reduced the share of subsequent questions from women by six points. The first question sets a template either way, so the only choice is whether to set it deliberately.
• The covert plant is standard practice. Ethics literature treats an undisclosed arranged question as deception, and disclosed variants work as well. In a small professional community where the same people attend every event, a discovered arrangement costs disproportionately more than it returns.

Where the evidence is genuinely strong

• Declared rehearsal plan — implementation intentions, Gollwitzer & Sheeran 2006: d = 0.65 across 94 tests and over 8,000 participants. Committing to how many run-throughs, of what type and when, outperforms knowing you should rehearse.
• Assertion-Evidence slides (Alley) — one of very few slide formats with comprehension and recall data behind it.
• Face-reading cannot carry “reading the room” — Barrett et al. 2019 and Savitsky & Gilovich 2003. Route diagnosis through pre-talk observation, one counted probe, and coarse behavioural signals.

False attributions to avoid

“Cold Open” is not a Presentation Patterns pattern — verified against the full glossary. Neither are Emergency Broadcast System or Dessert First. Cold Open is a borrowed screenwriting term. Talklet means roughly 20-minute units, not the 3-minute blocks it is often cited for. Lipsync (replaying a recorded tool interaction inside the talk) and Live on Tape (a recorded version of the whole talk, distributed) are two different patterns, frequently merged.

🔧 Tooling reality

Verified empirically on Marp CLI v4.4.1 / Core v4.3.1 rather than read from documentation:

• * and n) fragments emit data-marpit-fragment in HTML and flatten silently in PDF — every item lands on one page, no warning. Load-bearing reveals must become separate slides.
• A <video controls muted> slide exports to a PDF page whose only text is the control bar’s 0:00, with no image. poster= embeds that frame and removes the empty player.
• --pdf-notes attaches presenter notes as PDF annotations; --notes writes a separate text file; transition: is HTML-only.
• Native Mermaid is a Core v5 feature and unavailable on v4. Pre-render diagrams to SVG.

The practical consequence: if the PDF is your backup and the file the organiser receives, design to the PDF and treat HTML-only behaviour as decoration.

📖 Further reading

• Public Speaking — hub note
• TED Talks — Anderson: the idea is the payload; the five-step explanation ladder
• The Art of Public Speaking — Carnegie: emphasis, gesture, voice, and an argument-testing protocol that predates red-teaming by a century
• Marp, Marp CLI, Claude Code Marp Workflow — the deck toolchain
• Progressive Disclosure — the same layered-reveal principle applied to knowledge bases

---

Template: knowledge_note_info