PPT to Speech

Convert a slide deck into (1) a per-slide verbatim speech script, (2) speaker notes embedded back into the .pptx, and (3) a Q&A preparation section with anticipated questions and model answers.

The core insight: a good script is not a narration of what's on the slide. The audience can read. The script's job is to add what the slide cannot show — reasoning, transitions, emphasis, and story.

Pipeline

Four agents run in sequence. Each agent has its own spec file in agents/.

intake_agent  →  script_writer_agent  →  qa_prep_agent  →  delivery_agent

Agent handoff contracts are defined in shared/handoff_schemas.md. A missing required field triggers HANDOFF_INCOMPLETE and the pipeline stops.

Modes

Four modes, one per scenario. See MODE_REGISTRY.md for full trigger cues, oversight levels, and disambiguation rules.

⚠️ IRON RULE checkpoints

These checkpoints cannot be skipped or bypassed:

IR-1 — After intake_agent:
Present the Intake Record to the user. Do not call script_writer_agent until the user explicitly confirms. If they correct anything, update and re-present.

IR-2 — panel-discussion only, before script_writer_agent:
Confirm the panel theme and co-panelists' positions before writing anything. Missing this forces a full restart.

IR-3 — After script_writer_agent, if total runtime > 5% over time limit:
Show the word count arithmetic and ask the user to approve the overrun or trim before proceeding to qa_prep_agent.

Anti-patterns (enforced by script_writer_agent)

• No "This slide shows..." / "As you can see here..." — the audience can read
• No restating visible slide text — add the implication, not the repetition
• No uniform pacing — core contribution slides get 2–3× the time of housekeeping slides
• No unverified time estimates — word count must be counted and reconciled with stated minutes
• No missing transitions — every slide boundary needs an explicit bridge sentence
• No flattering Q&A questions — good prep questions hurt a little

Reference files