Write Talk Script

Write Talk Script

Produce a speaker script that fits an exact talk slot: every slide gets a word budget at the speaker's real pace, a cumulative timing mark, a written-out transition, and a marked cut list so running long on stage has a pre-planned answer. Slides tell the audience what to look at; this script is what the speaker actually says. It is the post-acceptance companion to make-slides.

When to use

• "Write the script for my 12-minute SIGSPATIAL talk" / "15-minute slot

including 3 minutes of Q&A — what do I say?"

• "Turn my deck into speaker notes I can rehearse from"
• "My talk is running long — what do I cut?"
• "I have a 5-minute lightning talk / spotlight"
• After make-slides, before the first timed rehearsal.

Inputs

1. The exact slot: total minutes, and whether Q&A is inside the slot or

handled separately. Ground truth is the presenter instructions and the acceptance email, not memory and not venue profiles.

1. The deck or an outline (slide titles in order). No deck yet? Run

make-slides first.

1. The paper (camera-ready preferred) — the only legitimate source for every

number and claim spoken.

1. Optional: venues/conferences/-.yml for venue context, and

the speaker's measured pace in words per minute.

Process

1. Pin down the slot facts. Ask for: total minutes, Q&A inside or

outside, session language, and whether the chair uses timecards. If a venue profile exists, use it for context only and re-verify against the live cfp_url / venue website — profiles do not encode talk slots, and slots change year to year. Typical shapes (verify, never assume): full papers 12–20 min including Q&A; CHI historically ~10 min + 5 Q&A; lightning/spotlight 1–7 min. Details in [references/timing-and-cuts.md](references/timing-and-cuts.md).

1. Calibrate the pace. Default to 130 wpm — a deliberate conference

pace. If the speaker can do the 1-minute read-aloud calibration (timing-and-cuts.md §2), use their measured number; suggest 110–120 wpm for non-native speakers or dense technical material. Never exceed 150.

1. Build the outline file. One line per slide (a - bullet or ##

heading), in deck order. Pin slides that need fixed time with a trailing | (e.g. - Key idea | 2). Title and takeaway slides typically pin at 0.5.

1. Compute the word budget. Run:

`` python3 scripts/word_budget.py --minutes 15 --qa 3 --outline outline.md \ [--wpm 120] [--venue-profile venues/conferences/.yml] ``

Also save the machine-readable form for step 6: python3 scripts/word_budget.py ... --json > budget.json. The script reserves a safety buffer (default 0.5 min) so the talk lands early, never at the bell. No outline yet? --slides N splits evenly.

1. Draft the script slide by slide, inside each slide's word budget, in

spoken register — short sentences, contractions, signposts, numbers written the way they are said. Follow [references/script-craft.md](references/script-craft.md) for the file format (one ## heading per slide, [bracketed] stage directions and timing marks, written-out transitions as the last sentence of each slide), opening and closing patterns, and figure walkthroughs. Every factual claim must trace to the paper — quote the paper's numbers, then render them speakable.

1. Time the draft. Run:

`` python3 scripts/script_timer.py talk-script.md --budget budget.json ``

(or --minutes 15 --qa 3 without a budget file). Fix OVER slides by cutting words, not by planning to speak faster; re-run until the verdict is WITHIN BUDGET. Exit codes: 0 within budget, 1 over, 2 bad input.

1. Build the cut list. Mark 2–3 cuttable passages totaling 10–15% of

the talk with ` … (each marker on its own line, no nesting, no crossing slide boundaries). What to cut first and what must never be cut: timing-and-cuts.md §6. Re-run script_timer.py` — it reports each cut's seconds saved and whether applying all cuts rescues an overrun.

1. Finalize the rehearsal package. Insert the budget's timing mark at

the top of each slide section (e.g. [4:30]), add the Q&A handoff line, and append the rehearsal protocol and the on-stage running-long protocol (timing-and-cuts.md §8–9). Tell the user the timing holds only after at least one full out-loud rehearsal with a timer.

Output

• talk-script.md — one ## section per slide: timing mark, word count,

spoken prose with stage directions, transition sentence, cut blocks; ends with the Q&A handoff and a rehearsal checklist.

• budget.json — the per-slide budget, reusable on every re-timing pass.
• A timing report (from script_timer.py) showing the talk fits, with the

cut list and seconds saved per cut.

Adapt to your discipline

The defaults target CS conference talks (12–20 min, slide-driven, English). For other fields, adjust the section shares in your outline (e.g. methods- heavy medical talks), the pace (--wpm), and the venue norms in timing-and-cuts.md — the scripts are field-agnostic.

Guardrails

• Never invent results, numbers, quotes, or citations for the script —

every claim spoken must come from the user's paper. New references go through verify-citations first.

• Timing is an estimate, not a guarantee: never tell the user the talk

"fits" without a real timed rehearsal; word-rate math ignores nerves, laser-pointer detours, and clicker failures (that is what the buffer and cut list are for).

• Venue profiles never override presenter instructions or the acceptance

email for slot length; re-verify against the live cfp_url before scripting.

• Preserve the speaker's voice and precise technical claims — spoken

register is a rewrite for the ear, not a simplification of the science, and not a read-aloud of paper paragraphs.

• Never upload the script or slides to any speaker portal, or submit

anything on the user's behalf.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ShaishavMaisuria
• Source: ShaishavMaisuria/research-paper-lifecycle-skills
• License: Apache-2.0
• Homepage: https://shaishavmaisuria.github.io/research-paper-lifecycle-skills/

Install and usage instructions live in the source repository linked above.