New deck

Builds one talk in this repo, end to end, in phases. The engine (navigation, presenter view, sync, print, lint) is finished and never changes; only src/slides/, src/deck.ts, src/assets/icons.ts, src/theme/ and docs/ do.

Before starting, read AGENTS.md and docs/authoring.md. See examples/crypto/docs/ for what the brief, plan, decisions and fact log look like when done. Open two or three slides in examples/wifi/slides/ and examples/crypto/slides/ to see the bar for slides and notes: script per step, Next line, acronyms, likely questions with answers, sources.

Stop at the end of every phase and wait for the speaker's review. Never carry on into the next phase in the same turn.

Where are we?

Check the state first and resume from the right phase:

• No docs/brief.md: Phase 0.
• Brief, but no docs/plan.md: Phase 1.
• Plan approved (docs/decisions.md has entries), and src/slides/ still holds the starter deck (data-id="idea" in 02-idea.html): Phase 2.
• Content written: Phase 3, then Phase 4.

Phase 0: Brief

1. Gather what the brief needs. If the speaker has already pasted an outline, a brief, lecture notes or files, read them first and only ask about what is missing. Ask in one round, with the host's question tool where available, or ordinary chat; offer options where clear-cut:
   • topic and working title; speaker name(s) exactly as shown; event or course; date
   • audience and what they already know; the goal of the talk
   • length, and whether a shorter version (second profile) is needed
   • sources: files (put them in docs/source/, never in the deck), links, required references
   • language and spelling
   • design direction: mood, colours, fonts, light or dark projector (or "decide later")
2. Copy docs/templates/brief.md to docs/brief.md and fill it. Draft the narrative and a slide plan yourself: one idea per slide, about 1 minute per slide on average, title first, references and questions last. Mark every draft line you invented, as opposed to what the speaker gave you, so they can see it.
3. Copy docs/templates/decisions.md and docs/templates/fact-log.md into docs/.
4. Set src/deck.ts id (kebab-case) and title, and package.json name and description. Replace the README's header (the kit's banner, badges and intro paragraphs, down to the sync GIF) with this talk's title and a short description, and keep the rest.

Stop. Show the brief's narrative and slide plan; ask for corrections.

Phase 1: Research and plan

If subagents are unavailable, run separate research and adversarial verification passes yourself, following the same evidence requirements. Disclose that verification was not independent. Use this fallback for Phase 3 as well.

1. List every factual claim the plan needs: numbers, dates, names, definitions, quotes.
2. Research with subagents in parallel, one per topic area, using available models. Each returns claims with primary sources (standards, papers, official docs, the speaker's own material first; vendor blogs and Wikipedia only as pointers), exact figures, dates and caveats.
3. Verify adversarially: one independent verification subagent per research batch tries to break each claim (outdated, misquoted, wrong unit, marketing figure, needs a qualifier). Keep only what survives; record disputes.
4. Write docs/fact-log.md (one id per claim) and docs/research/references.md.
5. Write docs/plan.md:
   • slide table: number, data-id, data-type, on-screen draft (about 30 words), build steps, target time, profiles
   • profiles and their total times; backup slides
   • diagrams to draw, described in one line each
   • numbered flags (C1, C2, …) wherever research contradicts or sharpens the brief, each with a recommendation and the fact-log ids behind it
   • structure questions (S1, S2, …): cuts, merges, exempt slides, anything else to decide

Stop. Ask the speaker to answer each flag ("ok" accepts the recommendation). Record the answers in docs/decisions.md, with date and source, before writing any slide.

Phase 2: Content

1. Remove the starter deck: delete src/slides/*.html. Write the new slides following docs/authoring.md exactly. Number the files in talk order (01-title.html, …).
2. Content comes from the brief as amended by docs/decisions.md, and nothing else. If a slide needs wording that was not approved, write it, mark the element with data-todo="why", and list it in your report.
3. Notes for every slide: Script with one block per step (spoken sentences, simple words), Next, Acronyms when there are any, If asked with 2 to 4 likely questions, and a <details> block with sources (fact-log ids), extra Q&A and timing. Use data-profile lines where a profile skips the following slide.
4. Diagrams: hand-drawn inline SVG, dgm-* classes. Add new classes to src/theme/diagrams.css with neutral styling that uses the existing tokens.
5. Icons: add what you use to src/assets/icons.ts. Before removing one, check it is not used by the engine (grep -r "'name'" src/engine, e.g. the timer's play, pause, rotate-ccw) or by an example (grep -r data-icon examples).
6. Update src/deck.ts: every profile, in order, with target times that add up to the brief.
7. Check, and fix until clean:
   • pnpm check (typecheck, Biome, unit tests, slide lint)
   • pnpm test:deck (every slide fits the canvas, axe, one PDF page per slide, offline)
   • pnpm shots, then look at every image in out/shots/: overflow, cramped text, empty space, a diagram that does not read. Fix layout problems with markup structure or theme classes, never with inline styles.

Stop. Report: slide list with times per profile, every data-todo, anything you could not fit, and the output of the checks (paste failures, do not summarise them away).

Phase 3: Verify

1. Adversarial fact check (verification subagent when available): every on-screen number and claim, and every "If asked" answer, against docs/fact-log.md and its sources. Report mismatches; fix only what is a transcription error, and flag the rest.
2. Copy check (copy-review subagent when available): the voice rules in docs/authoring.md, spelling consistency, and that each slide's Script covers what is on screen at each step.
3. Timing: the target times per profile add up to the brief's length, with about a minute spare for questions.
4. Build the rehearsal material: pnpm notes:pdf (and per profile if there are several).
5. Run pnpm test:e2e once, so the engine suite has run on this machine too.

Stop. Report each check with its real output.

Phase 4: Design

Run the retheme skill with the brief's design direction. Stop after it for review.

Then suggest ship-deck for the final build and the files to take to the room.

Notes

• Commit only when the speaker asks. Good commit points are the end of each phase.
• If the speaker wants to change content later, use add-slide or edit slides directly, and record content changes in docs/decisions.md.