Add: Presentation Talk Track Builder

[JagmeetChabraMSFT (jchha001)]

Add: Presentation Talk Track Builder

Adds a new skill that turns a slide deck (or slide images / pasted slide text)
into a first-person, ready-to-speak presenter script with delivery cues and a
calibrated timing table.

Why

Users often have a deck they need to present, but they either don't have
speaker notes at all or the existing notes are terse bullet fragments. This
skill turns whatever exists (slide content plus any existing notes, plus
optional supporting context and audience goals) into a natural spoken script
they can rehearse from or paste back into PowerPoint as speaker notes.

Highlights

• Notes are authoritative. When existing speaker notes contradict slide
  text, the script follows the notes and silently drops the contradicted
  slide text — that's what the author intended to say.
• First-person, spoken voice with [pause] cues, one-line slide-to-slide
  bridges, and a handoff line at the end.
• Timing calibration. Word budget = target minutes × chosen pace
  (deliberate 110-130 wpm, average 130-150 wpm, brisk 165-185 wpm) split
  across slides by note density; returns a per-slide word/time table with a
  total.
• Non-modification guardrail. Slide text, order, layouts, images, charts,
  and animations are never touched. If notes are written back into a .pptx,
  the helper writes them into a copied file only.
• Explicit inputs: Step 0 captures audience & goals; Step 1 asks for
  optional supporting context (product brief, prior version of the talk,
  FAQ, etc.) so the model has a real escape hatch when notes are thin.
• Confidentiality and PII guardrail covering enterprise sensitivity
  labels ("Confidential", "Internal Only", "Restricted"), unreleased
  figures, and customer/partner identifiers.

Platforms

Cowork, Copilot Studio, Scout — the bundled helper is Python stdlib only.

Local validation (per CONTRIBUTING.md)

Ran locally against the repo before opening this PR:

• npm run check:submissions → 13 submission(s) passed validation.
• npm run import:submissions → imported cleanly; produced
  src/content/skills/presentation-talk-track-builder.md and
  public/bundles/presentation-talk-track-builder.zip.
• npm run build → 57 pages built in ~34s, including
  /skills/presentation-talk-track-builder and new
  /tags/speaker-notes/ and /tags/pptx/ tag pages.

Files

submissions/presentation-talk-track-builder/
├── SKILL.md
├── metadata.json
├── scripts/pptx_talk_track.py
├── references/tone-examples.md
└── assets/example-notes.json

Sample output

Sample runs against real decks (COST AI Foundations for SALT, COST Virtual AI
Session 4) produced first-person spoken scripts with [pause] cues, natural
bridges between slides, and timing tables. On the notes-rich deck, the
"notes-are-authoritative" behaviour was exercised on slides where the visible
slide content is minimal (e.g. section dividers) — the script correctly
followed the speaker notes rather than the sparse slide text.

[JagmeetChabraMSFT (jchha001)]

@microsoft-github-policy-service agree company="Microsoft"

[adilei]

Reviewed and tested — stdlib-only Python (great sandbox fit for all three platforms), extract/template/apply subcommands round-trip a real deck cleanly, and the notes-are-authoritative design plus timing calibration are well thought out. CI green. Approving.